6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/register.hpp"
26 #include "ci/ciObjArray.hpp"
27 #include "ci/ciUtilities.hpp"
28 #include "classfile/javaClasses.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/c2/barrierSetC2.hpp"
32 #include "interpreter/interpreter.hpp"
33 #include "memory/resourceArea.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/graphKit.hpp"
38 #include "opto/idealKit.hpp"
39 #include "opto/intrinsicnode.hpp"
40 #include "opto/locknode.hpp"
41 #include "opto/machnode.hpp"
42 #include "opto/opaquenode.hpp"
43 #include "opto/parse.hpp"
44 #include "opto/rootnode.hpp"
45 #include "opto/runtime.hpp"
46 #include "opto/subtypenode.hpp"
47 #include "runtime/deoptimization.hpp"
48 #include "runtime/sharedRuntime.hpp"
49 #include "utilities/bitMap.inline.hpp"
50 #include "utilities/growableArray.hpp"
51 #include "utilities/powerOfTwo.hpp"
52
53 //----------------------------GraphKit-----------------------------------------
54 // Main utility constructor.
55 GraphKit::GraphKit(JVMState* jvms)
56 : Phase(Phase::Parser),
57 _env(C->env()),
58 _gvn(*C->initial_gvn()),
59 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
60 {
61 _exceptions = jvms->map()->next_exception();
62 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
63 set_jvms(jvms);
64 }
65
66 // Private constructor for parser.
67 GraphKit::GraphKit()
68 : Phase(Phase::Parser),
69 _env(C->env()),
70 _gvn(*C->initial_gvn()),
71 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
72 {
73 _exceptions = nullptr;
74 set_map(nullptr);
75 DEBUG_ONLY(_sp = -99);
76 DEBUG_ONLY(set_bci(-99));
77 }
78
79
80
81 //---------------------------clean_stack---------------------------------------
82 // Clear away rubbish from the stack area of the JVM state.
83 // This destroys any arguments that may be waiting on the stack.
328 }
329 static inline void add_one_req(Node* dstphi, Node* src) {
330 assert(is_hidden_merge(dstphi), "must be a special merge node");
331 assert(!is_hidden_merge(src), "must not be a special merge node");
332 dstphi->add_req(src);
333 }
334
335 //-----------------------combine_exception_states------------------------------
336 // This helper function combines exception states by building phis on a
337 // specially marked state-merging region. These regions and phis are
338 // untransformed, and can build up gradually. The region is marked by
339 // having a control input of its exception map, rather than null. Such
340 // regions do not appear except in this function, and in use_exception_state.
341 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
342 if (failing_internal()) {
343 return; // dying anyway...
344 }
345 JVMState* ex_jvms = ex_map->_jvms;
346 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
347 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
348 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
349 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
350 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
351 assert(ex_map->req() == phi_map->req(), "matching maps");
352 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
353 Node* hidden_merge_mark = root();
354 Node* region = phi_map->control();
355 MergeMemNode* phi_mem = phi_map->merged_memory();
356 MergeMemNode* ex_mem = ex_map->merged_memory();
357 if (region->in(0) != hidden_merge_mark) {
358 // The control input is not (yet) a specially-marked region in phi_map.
359 // Make it so, and build some phis.
360 region = new RegionNode(2);
361 _gvn.set_type(region, Type::CONTROL);
362 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
363 region->init_req(1, phi_map->control());
364 phi_map->set_control(region);
365 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
366 record_for_igvn(io_phi);
367 _gvn.set_type(io_phi, Type::ABIO);
368 phi_map->set_i_o(io_phi);
856 if (PrintMiscellaneous && (Verbose || WizardMode)) {
857 tty->print_cr("Zombie local %d: ", local);
858 jvms->dump();
859 }
860 return false;
861 }
862 }
863 }
864 return true;
865 }
866
867 #endif //ASSERT
868
869 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
870 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
871 ciMethod* cur_method = jvms->method();
872 int cur_bci = jvms->bci();
873 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
874 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
875 return Interpreter::bytecode_should_reexecute(code) ||
876 (is_anewarray && code == Bytecodes::_multianewarray);
877 // Reexecute _multianewarray bytecode which was replaced with
878 // sequence of [a]newarray. See Parse::do_multianewarray().
879 //
880 // Note: interpreter should not have it set since this optimization
881 // is limited by dimensions and guarded by flag so in some cases
882 // multianewarray() runtime calls will be generated and
883 // the bytecode should not be reexecutes (stack will not be reset).
884 } else {
885 return false;
886 }
887 }
888
889 // Helper function for adding JVMState and debug information to node
890 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
891 // Add the safepoint edges to the call (or other safepoint).
892
893 // Make sure dead locals are set to top. This
894 // should help register allocation time and cut down on the size
895 // of the deoptimization information.
896 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
947 }
948
949 // Presize the call:
950 DEBUG_ONLY(uint non_debug_edges = call->req());
951 call->add_req_batch(top(), youngest_jvms->debug_depth());
952 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
953
954 // Set up edges so that the call looks like this:
955 // Call [state:] ctl io mem fptr retadr
956 // [parms:] parm0 ... parmN
957 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
958 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
959 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
960 // Note that caller debug info precedes callee debug info.
961
962 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
963 uint debug_ptr = call->req();
964
965 // Loop over the map input edges associated with jvms, add them
966 // to the call node, & reset all offsets to match call node array.
967 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
968 uint debug_end = debug_ptr;
969 uint debug_start = debug_ptr - in_jvms->debug_size();
970 debug_ptr = debug_start; // back up the ptr
971
972 uint p = debug_start; // walks forward in [debug_start, debug_end)
973 uint j, k, l;
974 SafePointNode* in_map = in_jvms->map();
975 out_jvms->set_map(call);
976
977 if (can_prune_locals) {
978 assert(in_jvms->method() == out_jvms->method(), "sanity");
979 // If the current throw can reach an exception handler in this JVMS,
980 // then we must keep everything live that can reach that handler.
981 // As a quick and dirty approximation, we look for any handlers at all.
982 if (in_jvms->method()->has_exception_handlers()) {
983 can_prune_locals = false;
984 }
985 }
986
987 // Add the Locals
988 k = in_jvms->locoff();
989 l = in_jvms->loc_size();
990 out_jvms->set_locoff(p);
991 if (!can_prune_locals) {
992 for (j = 0; j < l; j++)
993 call->set_req(p++, in_map->in(k+j));
994 } else {
995 p += l; // already set to top above by add_req_batch
996 }
997
998 // Add the Expression Stack
999 k = in_jvms->stkoff();
1000 l = in_jvms->sp();
1001 out_jvms->set_stkoff(p);
1002 if (!can_prune_locals) {
1003 for (j = 0; j < l; j++)
1004 call->set_req(p++, in_map->in(k+j));
1005 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1006 // Divide stack into {S0,...,S1}, where S0 is set to top.
1007 uint s1 = stack_slots_not_pruned;
1008 stack_slots_not_pruned = 0; // for next iteration
1009 if (s1 > l) s1 = l;
1010 uint s0 = l - s1;
1011 p += s0; // skip the tops preinstalled by add_req_batch
1012 for (j = s0; j < l; j++)
1013 call->set_req(p++, in_map->in(k+j));
1014 } else {
1015 p += l; // already set to top above by add_req_batch
1016 }
1017
1018 // Add the Monitors
1019 k = in_jvms->monoff();
1020 l = in_jvms->mon_size();
1021 out_jvms->set_monoff(p);
1022 for (j = 0; j < l; j++)
1023 call->set_req(p++, in_map->in(k+j));
1024
1025 // Copy any scalar object fields.
1026 k = in_jvms->scloff();
1027 l = in_jvms->scl_size();
1028 out_jvms->set_scloff(p);
1029 for (j = 0; j < l; j++)
1030 call->set_req(p++, in_map->in(k+j));
1031
1032 // Finish the new jvms.
1033 out_jvms->set_endoff(p);
1034
1035 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1036 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1037 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1038 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1039 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1040 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1041
1042 // Update the two tail pointers in parallel.
1043 out_jvms = out_jvms->caller();
1044 in_jvms = in_jvms->caller();
1045 }
1046
1047 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1048
1049 // Test the correctness of JVMState::debug_xxx accessors:
1050 assert(call->jvms()->debug_start() == non_debug_edges, "");
1051 assert(call->jvms()->debug_end() == call->req(), "");
1052 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1053 }
1054
1055 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1056 Bytecodes::Code code = java_bc();
1057 if (code == Bytecodes::_wide) {
1058 code = method()->java_code_at_bci(bci() + 1);
1059 }
1060
1061 if (code != Bytecodes::_illegal) {
1062 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1198 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1199 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1200 return _gvn.transform( new AndLNode(conv, mask) );
1201 }
1202
1203 Node* GraphKit::ConvL2I(Node* offset) {
1204 // short-circuit a common case
1205 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1206 if (offset_con != (jlong)Type::OffsetBot) {
1207 return intcon((int) offset_con);
1208 }
1209 return _gvn.transform( new ConvL2INode(offset));
1210 }
1211
1212 //-------------------------load_object_klass-----------------------------------
1213 Node* GraphKit::load_object_klass(Node* obj) {
1214 // Special-case a fresh allocation to avoid building nodes:
1215 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1216 if (akls != nullptr) return akls;
1217 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1218 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1219 }
1220
1221 //-------------------------load_array_length-----------------------------------
1222 Node* GraphKit::load_array_length(Node* array) {
1223 // Special-case a fresh allocation to avoid building nodes:
1224 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1225 Node *alen;
1226 if (alloc == nullptr) {
1227 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1228 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1229 } else {
1230 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1231 }
1232 return alen;
1233 }
1234
1235 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1236 const TypeOopPtr* oop_type,
1237 bool replace_length_in_map) {
1238 Node* length = alloc->Ideal_length();
1247 replace_in_map(length, ccast);
1248 }
1249 return ccast;
1250 }
1251 }
1252 return length;
1253 }
1254
1255 //------------------------------do_null_check----------------------------------
1256 // Helper function to do a null pointer check. Returned value is
1257 // the incoming address with null casted away. You are allowed to use the
1258 // not-null value only if you are control dependent on the test.
1259 #ifndef PRODUCT
1260 extern uint explicit_null_checks_inserted,
1261 explicit_null_checks_elided;
1262 #endif
1263 Node* GraphKit::null_check_common(Node* value, BasicType type,
1264 // optional arguments for variations:
1265 bool assert_null,
1266 Node* *null_control,
1267 bool speculative) {
1268 assert(!assert_null || null_control == nullptr, "not both at once");
1269 if (stopped()) return top();
1270 NOT_PRODUCT(explicit_null_checks_inserted++);
1271
1272 // Construct null check
1273 Node *chk = nullptr;
1274 switch(type) {
1275 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1276 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1277 case T_ARRAY : // fall through
1278 type = T_OBJECT; // simplify further tests
1279 case T_OBJECT : {
1280 const Type *t = _gvn.type( value );
1281
1282 const TypeOopPtr* tp = t->isa_oopptr();
1283 if (tp != nullptr && !tp->is_loaded()
1284 // Only for do_null_check, not any of its siblings:
1285 && !assert_null && null_control == nullptr) {
1286 // Usually, any field access or invocation on an unloaded oop type
1287 // will simply fail to link, since the statically linked class is
1288 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1289 // the static class is loaded but the sharper oop type is not.
1290 // Rather than checking for this obscure case in lots of places,
1291 // we simply observe that a null check on an unloaded class
1355 }
1356 Node *oldcontrol = control();
1357 set_control(cfg);
1358 Node *res = cast_not_null(value);
1359 set_control(oldcontrol);
1360 NOT_PRODUCT(explicit_null_checks_elided++);
1361 return res;
1362 }
1363 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1364 if (cfg == nullptr) break; // Quit at region nodes
1365 depth++;
1366 }
1367 }
1368
1369 //-----------
1370 // Branch to failure if null
1371 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1372 Deoptimization::DeoptReason reason;
1373 if (assert_null) {
1374 reason = Deoptimization::reason_null_assert(speculative);
1375 } else if (type == T_OBJECT) {
1376 reason = Deoptimization::reason_null_check(speculative);
1377 } else {
1378 reason = Deoptimization::Reason_div0_check;
1379 }
1380 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1381 // ciMethodData::has_trap_at will return a conservative -1 if any
1382 // must-be-null assertion has failed. This could cause performance
1383 // problems for a method after its first do_null_assert failure.
1384 // Consider using 'Reason_class_check' instead?
1385
1386 // To cause an implicit null check, we set the not-null probability
1387 // to the maximum (PROB_MAX). For an explicit check the probability
1388 // is set to a smaller value.
1389 if (null_control != nullptr || too_many_traps(reason)) {
1390 // probability is less likely
1391 ok_prob = PROB_LIKELY_MAG(3);
1392 } else if (!assert_null &&
1393 (ImplicitNullCheckThreshold > 0) &&
1394 method() != nullptr &&
1395 (method()->method_data()->trap_count(reason)
1429 }
1430
1431 if (assert_null) {
1432 // Cast obj to null on this path.
1433 replace_in_map(value, zerocon(type));
1434 return zerocon(type);
1435 }
1436
1437 // Cast obj to not-null on this path, if there is no null_control.
1438 // (If there is a null_control, a non-null value may come back to haunt us.)
1439 if (type == T_OBJECT) {
1440 Node* cast = cast_not_null(value, false);
1441 if (null_control == nullptr || (*null_control) == top())
1442 replace_in_map(value, cast);
1443 value = cast;
1444 }
1445
1446 return value;
1447 }
1448
1449
1450 //------------------------------cast_not_null----------------------------------
1451 // Cast obj to not-null on this path
1452 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1453 const Type *t = _gvn.type(obj);
1454 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1455 // Object is already not-null?
1456 if( t == t_not_null ) return obj;
1457
1458 Node* cast = new CastPPNode(control(), obj,t_not_null);
1459 cast = _gvn.transform( cast );
1460
1461 // Scan for instances of 'obj' in the current JVM mapping.
1462 // These instances are known to be not-null after the test.
1463 if (do_replace_in_map)
1464 replace_in_map(obj, cast);
1465
1466 return cast; // Return casted value
1467 }
1468
1469 // Sometimes in intrinsics, we implicitly know an object is not null
1470 // (there's no actual null check) so we can cast it to not null. In
1471 // the course of optimizations, the input to the cast can become null.
1472 // In that case that data path will die and we need the control path
1473 // to become dead as well to keep the graph consistent. So we have to
1474 // add a check for null for which one branch can't be taken. It uses
1475 // an OpaqueNotNull node that will cause the check to be removed after loop
1476 // opts so the test goes away and the compiled code doesn't execute a
1477 // useless check.
1478 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1479 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1480 return value;
1481 }
1482 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1483 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1484 Node* opaq = _gvn.transform(new OpaqueNotNullNode(C, tst));
1485 IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1486 _gvn.set_type(iff, iff->Value(&_gvn));
1487 if (!tst->is_Con()) {
1488 record_for_igvn(iff);
1561 // These are layered on top of the factory methods in LoadNode and StoreNode,
1562 // and integrate with the parser's memory state and _gvn engine.
1563 //
1564
1565 // factory methods in "int adr_idx"
1566 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1567 MemNode::MemOrd mo,
1568 LoadNode::ControlDependency control_dependency,
1569 bool require_atomic_access,
1570 bool unaligned,
1571 bool mismatched,
1572 bool unsafe,
1573 uint8_t barrier_data) {
1574 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1575 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1576 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1577 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1578 Node* mem = memory(adr_idx);
1579 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1580 ld = _gvn.transform(ld);
1581 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1582 // Improve graph before escape analysis and boxing elimination.
1583 record_for_igvn(ld);
1584 if (ld->is_DecodeN()) {
1585 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1586 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1587 // a Phi). Recording such cases is still perfectly sound, but may be
1588 // unnecessary and result in some minor IGVN overhead.
1589 record_for_igvn(ld->in(1));
1590 }
1591 }
1592 return ld;
1593 }
1594
1595 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1596 MemNode::MemOrd mo,
1597 bool require_atomic_access,
1598 bool unaligned,
1599 bool mismatched,
1600 bool unsafe,
1614 if (unsafe) {
1615 st->as_Store()->set_unsafe_access();
1616 }
1617 st->as_Store()->set_barrier_data(barrier_data);
1618 st = _gvn.transform(st);
1619 set_memory(st, adr_idx);
1620 // Back-to-back stores can only remove intermediate store with DU info
1621 // so push on worklist for optimizer.
1622 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1623 record_for_igvn(st);
1624
1625 return st;
1626 }
1627
1628 Node* GraphKit::access_store_at(Node* obj,
1629 Node* adr,
1630 const TypePtr* adr_type,
1631 Node* val,
1632 const Type* val_type,
1633 BasicType bt,
1634 DecoratorSet decorators) {
1635 // Transformation of a value which could be null pointer (CastPP #null)
1636 // could be delayed during Parse (for example, in adjust_map_after_if()).
1637 // Execute transformation here to avoid barrier generation in such case.
1638 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1639 val = _gvn.makecon(TypePtr::NULL_PTR);
1640 }
1641
1642 if (stopped()) {
1643 return top(); // Dead path ?
1644 }
1645
1646 assert(val != nullptr, "not dead path");
1647
1648 C2AccessValuePtr addr(adr, adr_type);
1649 C2AccessValue value(val, val_type);
1650 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1651 if (access.is_raw()) {
1652 return _barrier_set->BarrierSetC2::store_at(access, value);
1653 } else {
1654 return _barrier_set->store_at(access, value);
1655 }
1656 }
1657
1658 Node* GraphKit::access_load_at(Node* obj, // containing obj
1659 Node* adr, // actual address to store val at
1660 const TypePtr* adr_type,
1661 const Type* val_type,
1662 BasicType bt,
1663 DecoratorSet decorators) {
1664 if (stopped()) {
1665 return top(); // Dead path ?
1666 }
1667
1668 C2AccessValuePtr addr(adr, adr_type);
1669 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1670 if (access.is_raw()) {
1671 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1672 } else {
1673 return _barrier_set->load_at(access, val_type);
1674 }
1675 }
1676
1677 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1678 const Type* val_type,
1679 BasicType bt,
1680 DecoratorSet decorators) {
1681 if (stopped()) {
1682 return top(); // Dead path ?
1683 }
1684
1685 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1686 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1687 if (access.is_raw()) {
1688 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1689 } else {
1754 Node* new_val,
1755 const Type* value_type,
1756 BasicType bt,
1757 DecoratorSet decorators) {
1758 C2AccessValuePtr addr(adr, adr_type);
1759 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1760 if (access.is_raw()) {
1761 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1762 } else {
1763 return _barrier_set->atomic_add_at(access, new_val, value_type);
1764 }
1765 }
1766
1767 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1768 return _barrier_set->clone(this, src, dst, size, is_array);
1769 }
1770
1771 //-------------------------array_element_address-------------------------
1772 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1773 const TypeInt* sizetype, Node* ctrl) {
1774 uint shift = exact_log2(type2aelembytes(elembt));
1775 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1776
1777 // short-circuit a common case (saves lots of confusing waste motion)
1778 jint idx_con = find_int_con(idx, -1);
1779 if (idx_con >= 0) {
1780 intptr_t offset = header + ((intptr_t)idx_con << shift);
1781 return basic_plus_adr(ary, offset);
1782 }
1783
1784 // must be correct type for alignment purposes
1785 Node* base = basic_plus_adr(ary, header);
1786 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1787 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1788 return basic_plus_adr(ary, base, scale);
1789 }
1790
1791 //-------------------------load_array_element-------------------------
1792 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1793 const Type* elemtype = arytype->elem();
1794 BasicType elembt = elemtype->array_element_basic_type();
1795 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1796 if (elembt == T_NARROWOOP) {
1797 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1798 }
1799 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1800 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1801 return ld;
1802 }
1803
1804 //-------------------------set_arguments_for_java_call-------------------------
1805 // Arguments (pre-popped from the stack) are taken from the JVMS.
1806 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1807 // Add the call arguments:
1808 uint nargs = call->method()->arg_size();
1809 for (uint i = 0; i < nargs; i++) {
1810 Node* arg = argument(i);
1811 call->init_req(i + TypeFunc::Parms, arg);
1812 }
1813 }
1814
1815 //---------------------------set_edges_for_java_call---------------------------
1816 // Connect a newly created call into the current JVMS.
1817 // A return value node (if any) is returned from set_edges_for_java_call.
1818 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1819
1820 // Add the predefined inputs:
1821 call->init_req( TypeFunc::Control, control() );
1822 call->init_req( TypeFunc::I_O , i_o() );
1823 call->init_req( TypeFunc::Memory , reset_memory() );
1824 call->init_req( TypeFunc::FramePtr, frameptr() );
1825 call->init_req( TypeFunc::ReturnAdr, top() );
1826
1827 add_safepoint_edges(call, must_throw);
1828
1829 Node* xcall = _gvn.transform(call);
1830
1831 if (xcall == top()) {
1832 set_control(top());
1833 return;
1834 }
1835 assert(xcall == call, "call identity is stable");
1836
1837 // Re-use the current map to produce the result.
1838
1839 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1840 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1841 set_all_memory_call(xcall, separate_io_proj);
1842
1843 //return xcall; // no need, caller already has it
1844 }
1845
1846 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1847 if (stopped()) return top(); // maybe the call folded up?
1848
1849 // Capture the return value, if any.
1850 Node* ret;
1851 if (call->method() == nullptr ||
1852 call->method()->return_type()->basic_type() == T_VOID)
1853 ret = top();
1854 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1855
1856 // Note: Since any out-of-line call can produce an exception,
1857 // we always insert an I_O projection from the call into the result.
1858
1859 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1860
1861 if (separate_io_proj) {
1862 // The caller requested separate projections be used by the fall
1863 // through and exceptional paths, so replace the projections for
1864 // the fall through path.
1865 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1866 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1867 }
1868 return ret;
1869 }
1870
1871 //--------------------set_predefined_input_for_runtime_call--------------------
1872 // Reading and setting the memory state is way conservative here.
1873 // The real problem is that I am not doing real Type analysis on memory,
1874 // so I cannot distinguish card mark stores from other stores. Across a GC
1875 // point the Store Barrier and the card mark memory has to agree. I cannot
1876 // have a card mark store and its barrier split across the GC point from
1877 // either above or below. Here I get that to happen by reading ALL of memory.
1878 // A better answer would be to separate out card marks from other memory.
1879 // For now, return the input memory state, so that it can be reused
1880 // after the call, if this call has restricted memory effects.
1881 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1882 // Set fixed predefined input arguments
1883 call->init_req(TypeFunc::Control, control());
1884 call->init_req(TypeFunc::I_O, top()); // does no i/o
1885 call->init_req(TypeFunc::ReturnAdr, top());
1886 if (call->is_CallLeafPure()) {
1887 call->init_req(TypeFunc::Memory, top());
1949 if (use->is_MergeMem()) {
1950 wl.push(use);
1951 }
1952 }
1953 }
1954
1955 // Replace the call with the current state of the kit.
1956 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
1957 JVMState* ejvms = nullptr;
1958 if (has_exceptions()) {
1959 ejvms = transfer_exceptions_into_jvms();
1960 }
1961
1962 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1963 ReplacedNodes replaced_nodes_exception;
1964 Node* ex_ctl = top();
1965
1966 SafePointNode* final_state = stop();
1967
1968 // Find all the needed outputs of this call
1969 CallProjections callprojs;
1970 call->extract_projections(&callprojs, true, do_asserts);
1971
1972 Unique_Node_List wl;
1973 Node* init_mem = call->in(TypeFunc::Memory);
1974 Node* final_mem = final_state->in(TypeFunc::Memory);
1975 Node* final_ctl = final_state->in(TypeFunc::Control);
1976 Node* final_io = final_state->in(TypeFunc::I_O);
1977
1978 // Replace all the old call edges with the edges from the inlining result
1979 if (callprojs.fallthrough_catchproj != nullptr) {
1980 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1981 }
1982 if (callprojs.fallthrough_memproj != nullptr) {
1983 if (final_mem->is_MergeMem()) {
1984 // Parser's exits MergeMem was not transformed but may be optimized
1985 final_mem = _gvn.transform(final_mem);
1986 }
1987 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1988 add_mergemem_users_to_worklist(wl, final_mem);
1989 }
1990 if (callprojs.fallthrough_ioproj != nullptr) {
1991 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1992 }
1993
1994 // Replace the result with the new result if it exists and is used
1995 if (callprojs.resproj != nullptr && result != nullptr) {
1996 C->gvn_replace_by(callprojs.resproj, result);
1997 }
1998
1999 if (ejvms == nullptr) {
2000 // No exception edges to simply kill off those paths
2001 if (callprojs.catchall_catchproj != nullptr) {
2002 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
2003 }
2004 if (callprojs.catchall_memproj != nullptr) {
2005 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
2006 }
2007 if (callprojs.catchall_ioproj != nullptr) {
2008 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
2009 }
2010 // Replace the old exception object with top
2011 if (callprojs.exobj != nullptr) {
2012 C->gvn_replace_by(callprojs.exobj, C->top());
2013 }
2014 } else {
2015 GraphKit ekit(ejvms);
2016
2017 // Load my combined exception state into the kit, with all phis transformed:
2018 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2019 replaced_nodes_exception = ex_map->replaced_nodes();
2020
2021 Node* ex_oop = ekit.use_exception_state(ex_map);
2022
2023 if (callprojs.catchall_catchproj != nullptr) {
2024 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
2025 ex_ctl = ekit.control();
2026 }
2027 if (callprojs.catchall_memproj != nullptr) {
2028 Node* ex_mem = ekit.reset_memory();
2029 C->gvn_replace_by(callprojs.catchall_memproj, ex_mem);
2030 add_mergemem_users_to_worklist(wl, ex_mem);
2031 }
2032 if (callprojs.catchall_ioproj != nullptr) {
2033 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
2034 }
2035
2036 // Replace the old exception object with the newly created one
2037 if (callprojs.exobj != nullptr) {
2038 C->gvn_replace_by(callprojs.exobj, ex_oop);
2039 }
2040 }
2041
2042 // Disconnect the call from the graph
2043 call->disconnect_inputs(C);
2044 C->gvn_replace_by(call, C->top());
2045
2046 // Clean up any MergeMems that feed other MergeMems since the
2047 // optimizer doesn't like that.
2048 while (wl.size() > 0) {
2049 _gvn.transform(wl.pop());
2050 }
2051
2052 if (callprojs.fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2053 replaced_nodes.apply(C, final_ctl);
2054 }
2055 if (!ex_ctl->is_top() && do_replaced_nodes) {
2056 replaced_nodes_exception.apply(C, ex_ctl);
2057 }
2058 }
2059
2060
2061 //------------------------------increment_counter------------------------------
2062 // for statistics: increment a VM counter by 1
2063
2064 void GraphKit::increment_counter(address counter_addr) {
2065 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2066 increment_counter(adr1);
2067 }
2068
2069 void GraphKit::increment_counter(Node* counter_addr) {
2070 Node* ctrl = control();
2071 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2072 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2232 *
2233 * @param n node that the type applies to
2234 * @param exact_kls type from profiling
2235 * @param maybe_null did profiling see null?
2236 *
2237 * @return node with improved type
2238 */
2239 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2240 const Type* current_type = _gvn.type(n);
2241 assert(UseTypeSpeculation, "type speculation must be on");
2242
2243 const TypePtr* speculative = current_type->speculative();
2244
2245 // Should the klass from the profile be recorded in the speculative type?
2246 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2247 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2248 const TypeOopPtr* xtype = tklass->as_instance_type();
2249 assert(xtype->klass_is_exact(), "Should be exact");
2250 // Any reason to believe n is not null (from this profiling or a previous one)?
2251 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2252 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2253 // record the new speculative type's depth
2254 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2255 speculative = speculative->with_inline_depth(jvms()->depth());
2256 } else if (current_type->would_improve_ptr(ptr_kind)) {
2257 // Profiling report that null was never seen so we can change the
2258 // speculative type to non null ptr.
2259 if (ptr_kind == ProfileAlwaysNull) {
2260 speculative = TypePtr::NULL_PTR;
2261 } else {
2262 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2263 const TypePtr* ptr = TypePtr::NOTNULL;
2264 if (speculative != nullptr) {
2265 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2266 } else {
2267 speculative = ptr;
2268 }
2269 }
2270 }
2271
2272 if (speculative != current_type->speculative()) {
2273 // Build a type with a speculative type (what we think we know
2274 // about the type but will need a guard when we use it)
2275 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2276 // We're changing the type, we need a new CheckCast node to carry
2277 // the new type. The new type depends on the control: what
2278 // profiling tells us is only valid from here as far as we can
2279 // tell.
2280 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2281 cast = _gvn.transform(cast);
2282 replace_in_map(n, cast);
2283 n = cast;
2284 }
2285
2286 return n;
2287 }
2288
2289 /**
2290 * Record profiling data from receiver profiling at an invoke with the
2291 * type system so that it can propagate it (speculation)
2292 *
2293 * @param n receiver node
2294 *
2295 * @return node with improved type
2296 */
2297 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2298 if (!UseTypeSpeculation) {
2299 return n;
2300 }
2301 ciKlass* exact_kls = profile_has_unique_klass();
2302 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2303 if ((java_bc() == Bytecodes::_checkcast ||
2304 java_bc() == Bytecodes::_instanceof ||
2305 java_bc() == Bytecodes::_aastore) &&
2306 method()->method_data()->is_mature()) {
2307 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2308 if (data != nullptr) {
2309 if (!data->as_BitData()->null_seen()) {
2310 ptr_kind = ProfileNeverNull;
2311 } else {
2312 if (TypeProfileCasts) {
2313 assert(data->is_ReceiverTypeData(), "bad profile data type");
2314 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2315 uint i = 0;
2316 for (; i < call->row_limit(); i++) {
2317 ciKlass* receiver = call->receiver(i);
2318 if (receiver != nullptr) {
2319 break;
2320 }
2321 }
2322 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2323 }
2324 }
2325 }
2326 }
2327 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2328 }
2329
2330 /**
2331 * Record profiling data from argument profiling at an invoke with the
2332 * type system so that it can propagate it (speculation)
2333 *
2334 * @param dest_method target method for the call
2335 * @param bc what invoke bytecode is this?
2336 */
2337 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2338 if (!UseTypeSpeculation) {
2339 return;
2340 }
2341 const TypeFunc* tf = TypeFunc::make(dest_method);
2342 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2343 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2344 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2345 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2346 if (is_reference_type(targ->basic_type())) {
2347 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2348 ciKlass* better_type = nullptr;
2349 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2350 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2351 }
2352 i++;
2353 }
2354 }
2355 }
2356
2357 /**
2358 * Record profiling data from parameter profiling at an invoke with
2359 * the type system so that it can propagate it (speculation)
2360 */
2361 void GraphKit::record_profiled_parameters_for_speculation() {
2362 if (!UseTypeSpeculation) {
2363 return;
2364 }
2365 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2485 // The first null ends the list.
2486 Node* parm0, Node* parm1,
2487 Node* parm2, Node* parm3,
2488 Node* parm4, Node* parm5,
2489 Node* parm6, Node* parm7) {
2490 assert(call_addr != nullptr, "must not call null targets");
2491
2492 // Slow-path call
2493 bool is_leaf = !(flags & RC_NO_LEAF);
2494 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2495 if (call_name == nullptr) {
2496 assert(!is_leaf, "must supply name for leaf");
2497 call_name = OptoRuntime::stub_name(call_addr);
2498 }
2499 CallNode* call;
2500 if (!is_leaf) {
2501 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2502 } else if (flags & RC_NO_FP) {
2503 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2504 } else if (flags & RC_VECTOR){
2505 uint num_bits = call_type->range()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2506 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2507 } else if (flags & RC_PURE) {
2508 call = new CallLeafPureNode(call_type, call_addr, call_name, adr_type);
2509 } else {
2510 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2511 }
2512
2513 // The following is similar to set_edges_for_java_call,
2514 // except that the memory effects of the call are restricted to AliasIdxRaw.
2515
2516 // Slow path call has no side-effects, uses few values
2517 bool wide_in = !(flags & RC_NARROW_MEM);
2518 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2519
2520 Node* prev_mem = nullptr;
2521 if (wide_in) {
2522 prev_mem = set_predefined_input_for_runtime_call(call);
2523 } else {
2524 assert(!wide_out, "narrow in => narrow out");
2525 Node* narrow_mem = memory(adr_type);
2526 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2527 }
2528
2529 // Hook each parm in order. Stop looking at the first null.
2530 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2531 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2532 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2533 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2534 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2535 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2536 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2537 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2538 /* close each nested if ===> */ } } } } } } } }
2539 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
2540
2541 if (!is_leaf) {
2542 // Non-leaves can block and take safepoints:
2543 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2544 }
2545 // Non-leaves can throw exceptions:
2546 if (has_io) {
2547 call->set_req(TypeFunc::I_O, i_o());
2548 }
2549
2550 if (flags & RC_UNCOMMON) {
2551 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2552 // (An "if" probability corresponds roughly to an unconditional count.
2553 // Sort of.)
2554 call->set_cnt(PROB_UNLIKELY_MAG(4));
2555 }
2556
2557 Node* c = _gvn.transform(call);
2558 assert(c == call, "cannot disappear");
2559
2567
2568 if (has_io) {
2569 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2570 }
2571 return call;
2572
2573 }
2574
2575 // i2b
2576 Node* GraphKit::sign_extend_byte(Node* in) {
2577 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2578 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2579 }
2580
2581 // i2s
2582 Node* GraphKit::sign_extend_short(Node* in) {
2583 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2584 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2585 }
2586
2587 //------------------------------merge_memory-----------------------------------
2588 // Merge memory from one path into the current memory state.
2589 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2590 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2591 Node* old_slice = mms.force_memory();
2592 Node* new_slice = mms.memory2();
2593 if (old_slice != new_slice) {
2594 PhiNode* phi;
2595 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2596 if (mms.is_empty()) {
2597 // clone base memory Phi's inputs for this memory slice
2598 assert(old_slice == mms.base_memory(), "sanity");
2599 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2600 _gvn.set_type(phi, Type::MEMORY);
2601 for (uint i = 1; i < phi->req(); i++) {
2602 phi->init_req(i, old_slice->in(i));
2603 }
2604 } else {
2605 phi = old_slice->as_Phi(); // Phi was generated already
2606 }
2663 gvn.transform(iff);
2664 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2665 return iff;
2666 }
2667
2668 //-------------------------------gen_subtype_check-----------------------------
2669 // Generate a subtyping check. Takes as input the subtype and supertype.
2670 // Returns 2 values: sets the default control() to the true path and returns
2671 // the false path. Only reads invariant memory; sets no (visible) memory.
2672 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2673 // but that's not exposed to the optimizer. This call also doesn't take in an
2674 // Object; if you wish to check an Object you need to load the Object's class
2675 // prior to coming here.
2676 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
2677 ciMethod* method, int bci) {
2678 Compile* C = gvn.C;
2679 if ((*ctrl)->is_top()) {
2680 return C->top();
2681 }
2682
2683 // Fast check for identical types, perhaps identical constants.
2684 // The types can even be identical non-constants, in cases
2685 // involving Array.newInstance, Object.clone, etc.
2686 if (subklass == superklass)
2687 return C->top(); // false path is dead; no test needed.
2688
2689 if (gvn.type(superklass)->singleton()) {
2690 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2691 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
2692
2693 // In the common case of an exact superklass, try to fold up the
2694 // test before generating code. You may ask, why not just generate
2695 // the code and then let it fold up? The answer is that the generated
2696 // code will necessarily include null checks, which do not always
2697 // completely fold away. If they are also needless, then they turn
2698 // into a performance loss. Example:
2699 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2700 // Here, the type of 'fa' is often exact, so the store check
2701 // of fa[1]=x will fold up, without testing the nullness of x.
2702 //
2703 // At macro expansion, we would have already folded the SubTypeCheckNode
2704 // being expanded here because we always perform the static sub type
2705 // check in SubTypeCheckNode::sub() regardless of whether
2706 // StressReflectiveCode is set or not. We can therefore skip this
2707 // static check when StressReflectiveCode is on.
2708 switch (C->static_subtype_check(superk, subk)) {
2709 case Compile::SSC_always_false:
2710 {
2711 Node* always_fail = *ctrl;
2712 *ctrl = gvn.C->top();
2713 return always_fail;
2714 }
2715 case Compile::SSC_always_true:
2716 return C->top();
2717 case Compile::SSC_easy_test:
2718 {
2719 // Just do a direct pointer compare and be done.
2720 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2721 *ctrl = gvn.transform(new IfTrueNode(iff));
2722 return gvn.transform(new IfFalseNode(iff));
2723 }
2724 case Compile::SSC_full_test:
2725 break;
2726 default:
2727 ShouldNotReachHere();
2728 }
2729 }
2730
2731 // %%% Possible further optimization: Even if the superklass is not exact,
2732 // if the subklass is the unique subtype of the superklass, the check
2733 // will always succeed. We could leave a dependency behind to ensure this.
2734
2735 // First load the super-klass's check-offset
2736 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2737 Node* m = C->immutable_memory();
2738 Node *chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2739 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2740 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
2778 gvn.record_for_igvn(r_ok_subtype);
2779
2780 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
2781 // SubTypeCheck node
2782 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
2783 ciCallProfile profile = method->call_profile_at_bci(bci);
2784 float total_prob = 0;
2785 for (int i = 0; profile.has_receiver(i); ++i) {
2786 float prob = profile.receiver_prob(i);
2787 total_prob += prob;
2788 }
2789 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
2790 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2791 for (int i = 0; profile.has_receiver(i); ++i) {
2792 ciKlass* klass = profile.receiver(i);
2793 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
2794 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
2795 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
2796 continue;
2797 }
2798 float prob = profile.receiver_prob(i);
2799 ConNode* klass_node = gvn.makecon(klass_t);
2800 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
2801 Node* iftrue = gvn.transform(new IfTrueNode(iff));
2802
2803 if (result == Compile::SSC_always_true) {
2804 r_ok_subtype->add_req(iftrue);
2805 } else {
2806 assert(result == Compile::SSC_always_false, "");
2807 r_not_subtype->add_req(iftrue);
2808 }
2809 *ctrl = gvn.transform(new IfFalseNode(iff));
2810 }
2811 }
2812 }
2813
2814 // See if we get an immediate positive hit. Happens roughly 83% of the
2815 // time. Test to see if the value loaded just previously from the subklass
2816 // is exactly the superklass.
2817 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2831 igvn->remove_globally_dead_node(r_not_subtype);
2832 }
2833 return not_subtype_ctrl;
2834 }
2835
2836 r_ok_subtype->init_req(1, iftrue1);
2837
2838 // Check for immediate negative hit. Happens roughly 11% of the time (which
2839 // is roughly 63% of the remaining cases). Test to see if the loaded
2840 // check-offset points into the subklass display list or the 1-element
2841 // cache. If it points to the display (and NOT the cache) and the display
2842 // missed then it's not a subtype.
2843 Node *cacheoff = gvn.intcon(cacheoff_con);
2844 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2845 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
2846 *ctrl = gvn.transform(new IfFalseNode(iff2));
2847
2848 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2849 // No performance impact (too rare) but allows sharing of secondary arrays
2850 // which has some footprint reduction.
2851 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2852 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
2853 *ctrl = gvn.transform(new IfFalseNode(iff3));
2854
2855 // -- Roads not taken here: --
2856 // We could also have chosen to perform the self-check at the beginning
2857 // of this code sequence, as the assembler does. This would not pay off
2858 // the same way, since the optimizer, unlike the assembler, can perform
2859 // static type analysis to fold away many successful self-checks.
2860 // Non-foldable self checks work better here in second position, because
2861 // the initial primary superclass check subsumes a self-check for most
2862 // types. An exception would be a secondary type like array-of-interface,
2863 // which does not appear in its own primary supertype display.
2864 // Finally, we could have chosen to move the self-check into the
2865 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2866 // dependent manner. But it is worthwhile to have the check here,
2867 // where it can be perhaps be optimized. The cost in code space is
2868 // small (register compare, branch).
2869
2870 // Now do a linear scan of the secondary super-klass array. Again, no real
2871 // performance impact (too rare) but it's gotta be done.
2872 // Since the code is rarely used, there is no penalty for moving it
2873 // out of line, and it can only improve I-cache density.
2874 // The decision to inline or out-of-line this final check is platform
2875 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2876 Node* psc = gvn.transform(
2877 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2878
2879 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2880 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
2881 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
2882
2883 // Return false path; set default control to true path.
2884 *ctrl = gvn.transform(r_ok_subtype);
2885 return gvn.transform(r_not_subtype);
2886 }
2887
2888 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
2889 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
2890 if (expand_subtype_check) {
2891 MergeMemNode* mem = merged_memory();
2892 Node* ctrl = control();
2893 Node* subklass = obj_or_subklass;
2894 if (!_gvn.type(obj_or_subklass)->isa_klassptr()) {
2895 subklass = load_object_klass(obj_or_subklass);
2896 }
2897
2898 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
2899 set_control(ctrl);
2900 return n;
2901 }
2902
2903 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
2904 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
2905 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2906 set_control(_gvn.transform(new IfTrueNode(iff)));
2907 return _gvn.transform(new IfFalseNode(iff));
2908 }
2909
2910 // Profile-driven exact type check:
2911 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2912 float prob,
2913 Node* *casted_receiver) {
2914 assert(!klass->is_interface(), "no exact type check on interfaces");
2915
2916 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
2917 Node* recv_klass = load_object_klass(receiver);
2918 Node* want_klass = makecon(tklass);
2919 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
2920 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
2921 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2922 set_control( _gvn.transform(new IfTrueNode (iff)));
2923 Node* fail = _gvn.transform(new IfFalseNode(iff));
2924
2925 if (!stopped()) {
2926 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
2927 const TypeOopPtr* recvx_type = tklass->as_instance_type();
2928 assert(recvx_type->klass_is_exact(), "");
2929
2930 if (!receiver_type->higher_equal(recvx_type)) { // ignore redundant casts
2931 // Subsume downstream occurrences of receiver with a cast to
2932 // recv_xtype, since now we know what the type will be.
2933 Node* cast = new CheckCastPPNode(control(), receiver, recvx_type);
2934 (*casted_receiver) = _gvn.transform(cast);
2935 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
2936 // (User must make the replace_in_map call.)
2937 }
2938 }
2939
2940 return fail;
2941 }
2942
2943 //------------------------------subtype_check_receiver-------------------------
2944 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
2945 Node** casted_receiver) {
2946 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
2947 Node* want_klass = makecon(tklass);
2948
2949 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
2950
2951 // Ignore interface type information until interface types are properly tracked.
2952 if (!stopped() && !klass->is_interface()) {
2953 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
2954 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
2955 if (!receiver_type->higher_equal(recv_type)) { // ignore redundant casts
2956 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
2957 (*casted_receiver) = _gvn.transform(cast);
2958 }
2959 }
2960
2961 return slow_ctl;
2962 }
2963
2964 //------------------------------seems_never_null-------------------------------
2965 // Use null_seen information if it is available from the profile.
2966 // If we see an unexpected null at a type check we record it and force a
2967 // recompile; the offending check will be recompiled to handle nulls.
2968 // If we see several offending BCIs, then all checks in the
2969 // method will be recompiled.
2970 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2971 speculating = !_gvn.type(obj)->speculative_maybe_null();
2972 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2973 if (UncommonNullCast // Cutout for this technique
2974 && obj != null() // And not the -Xcomp stupid case?
2975 && !too_many_traps(reason)
2976 ) {
2977 if (speculating) {
3046
3047 //------------------------maybe_cast_profiled_receiver-------------------------
3048 // If the profile has seen exactly one type, narrow to exactly that type.
3049 // Subsequent type checks will always fold up.
3050 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3051 const TypeKlassPtr* require_klass,
3052 ciKlass* spec_klass,
3053 bool safe_for_replace) {
3054 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3055
3056 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3057
3058 // Make sure we haven't already deoptimized from this tactic.
3059 if (too_many_traps_or_recompiles(reason))
3060 return nullptr;
3061
3062 // (No, this isn't a call, but it's enough like a virtual call
3063 // to use the same ciMethod accessor to get the profile info...)
3064 // If we have a speculative type use it instead of profiling (which
3065 // may not help us)
3066 ciKlass* exact_kls = spec_klass == nullptr ? profile_has_unique_klass() : spec_klass;
3067 if (exact_kls != nullptr) {// no cast failures here
3068 if (require_klass == nullptr ||
3069 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3070 // If we narrow the type to match what the type profile sees or
3071 // the speculative type, we can then remove the rest of the
3072 // cast.
3073 // This is a win, even if the exact_kls is very specific,
3074 // because downstream operations, such as method calls,
3075 // will often benefit from the sharper type.
3076 Node* exact_obj = not_null_obj; // will get updated in place...
3077 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3078 &exact_obj);
3079 { PreserveJVMState pjvms(this);
3080 set_control(slow_ctl);
3081 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3082 }
3083 if (safe_for_replace) {
3084 replace_in_map(not_null_obj, exact_obj);
3085 }
3086 return exact_obj;
3176 // If not_null_obj is dead, only null-path is taken
3177 if (stopped()) { // Doing instance-of on a null?
3178 set_control(null_ctl);
3179 return intcon(0);
3180 }
3181 region->init_req(_null_path, null_ctl);
3182 phi ->init_req(_null_path, intcon(0)); // Set null path value
3183 if (null_ctl == top()) {
3184 // Do this eagerly, so that pattern matches like is_diamond_phi
3185 // will work even during parsing.
3186 assert(_null_path == PATH_LIMIT-1, "delete last");
3187 region->del_req(_null_path);
3188 phi ->del_req(_null_path);
3189 }
3190
3191 // Do we know the type check always succeed?
3192 bool known_statically = false;
3193 if (_gvn.type(superklass)->singleton()) {
3194 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3195 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3196 if (subk->is_loaded()) {
3197 int static_res = C->static_subtype_check(superk, subk);
3198 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3199 }
3200 }
3201
3202 if (!known_statically) {
3203 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3204 // We may not have profiling here or it may not help us. If we
3205 // have a speculative type use it to perform an exact cast.
3206 ciKlass* spec_obj_type = obj_type->speculative_type();
3207 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3208 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3209 if (stopped()) { // Profile disagrees with this path.
3210 set_control(null_ctl); // Null is the only remaining possibility.
3211 return intcon(0);
3212 }
3213 if (cast_obj != nullptr) {
3214 not_null_obj = cast_obj;
3215 }
3216 }
3232 record_for_igvn(region);
3233
3234 // If we know the type check always succeeds then we don't use the
3235 // profiling data at this bytecode. Don't lose it, feed it to the
3236 // type system as a speculative type.
3237 if (safe_for_replace) {
3238 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3239 replace_in_map(obj, casted_obj);
3240 }
3241
3242 return _gvn.transform(phi);
3243 }
3244
3245 //-------------------------------gen_checkcast---------------------------------
3246 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3247 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3248 // uncommon-trap paths work. Adjust stack after this call.
3249 // If failure_control is supplied and not null, it is filled in with
3250 // the control edge for the cast failure. Otherwise, an appropriate
3251 // uncommon trap or exception is thrown.
3252 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3253 Node* *failure_control) {
3254 kill_dead_locals(); // Benefit all the uncommon traps
3255 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3256 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3257 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3258
3259 // Fast cutout: Check the case that the cast is vacuously true.
3260 // This detects the common cases where the test will short-circuit
3261 // away completely. We do this before we perform the null check,
3262 // because if the test is going to turn into zero code, we don't
3263 // want a residual null check left around. (Causes a slowdown,
3264 // for example, in some objArray manipulations, such as a[i]=a[j].)
3265 if (improved_klass_ptr_type->singleton()) {
3266 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3267 if (objtp != nullptr) {
3268 switch (C->static_subtype_check(improved_klass_ptr_type, objtp->as_klass_type())) {
3269 case Compile::SSC_always_true:
3270 // If we know the type check always succeed then we don't use
3271 // the profiling data at this bytecode. Don't lose it, feed it
3272 // to the type system as a speculative type.
3273 return record_profiled_receiver_for_speculation(obj);
3274 case Compile::SSC_always_false:
3275 // It needs a null check because a null will *pass* the cast check.
3276 // A non-null value will always produce an exception.
3277 if (!objtp->maybe_null()) {
3278 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3279 Deoptimization::DeoptReason reason = is_aastore ?
3280 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3281 builtin_throw(reason);
3282 return top();
3283 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3284 return null_assert(obj);
3285 }
3286 break; // Fall through to full check
3287 default:
3288 break;
3289 }
3290 }
3291 }
3292
3293 ciProfileData* data = nullptr;
3294 bool safe_for_replace = false;
3295 if (failure_control == nullptr) { // use MDO in regular case only
3296 assert(java_bc() == Bytecodes::_aastore ||
3297 java_bc() == Bytecodes::_checkcast,
3298 "interpreter profiles type checks only for these BCs");
3299 data = method()->method_data()->bci_to_data(bci());
3300 safe_for_replace = true;
3301 }
3302
3303 // Make the merge point
3304 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3305 RegionNode* region = new RegionNode(PATH_LIMIT);
3306 Node* phi = new PhiNode(region, toop);
3307 C->set_has_split_ifs(true); // Has chance for split-if optimization
3308
3309 // Use null-cast information if it is available
3310 bool speculative_not_null = false;
3311 bool never_see_null = ((failure_control == nullptr) // regular case only
3312 && seems_never_null(obj, data, speculative_not_null));
3313
3314 // Null check; get casted pointer; set region slot 3
3315 Node* null_ctl = top();
3316 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3317
3318 // If not_null_obj is dead, only null-path is taken
3319 if (stopped()) { // Doing instance-of on a null?
3320 set_control(null_ctl);
3321 return null();
3322 }
3323 region->init_req(_null_path, null_ctl);
3324 phi ->init_req(_null_path, null()); // Set null path value
3325 if (null_ctl == top()) {
3326 // Do this eagerly, so that pattern matches like is_diamond_phi
3327 // will work even during parsing.
3328 assert(_null_path == PATH_LIMIT-1, "delete last");
3329 region->del_req(_null_path);
3330 phi ->del_req(_null_path);
3331 }
3332
3333 Node* cast_obj = nullptr;
3334 if (improved_klass_ptr_type->klass_is_exact()) {
3335 // The following optimization tries to statically cast the speculative type of the object
3336 // (for example obtained during profiling) to the type of the superklass and then do a
3337 // dynamic check that the type of the object is what we expect. To work correctly
3338 // for checkcast and aastore the type of superklass should be exact.
3339 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3340 // We may not have profiling here or it may not help us. If we have
3341 // a speculative type use it to perform an exact cast.
3342 ciKlass* spec_obj_type = obj_type->speculative_type();
3343 if (spec_obj_type != nullptr || data != nullptr) {
3344 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3345 if (cast_obj != nullptr) {
3346 if (failure_control != nullptr) // failure is now impossible
3347 (*failure_control) = top();
3348 // adjust the type of the phi to the exact klass:
3349 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3350 }
3351 }
3352 }
3353
3354 if (cast_obj == nullptr) {
3355 // Generate the subtype check
3356 Node* improved_superklass = superklass;
3357 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3358 improved_superklass = makecon(improved_klass_ptr_type);
3359 }
3360 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3361
3362 // Plug in success path into the merge
3363 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3364 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3365 if (failure_control == nullptr) {
3366 if (not_subtype_ctrl != top()) { // If failure is possible
3367 PreserveJVMState pjvms(this);
3368 set_control(not_subtype_ctrl);
3369 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3370 Deoptimization::DeoptReason reason = is_aastore ?
3371 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3372 builtin_throw(reason);
3373 }
3374 } else {
3375 (*failure_control) = not_subtype_ctrl;
3376 }
3377 }
3378
3379 region->init_req(_obj_path, control());
3380 phi ->init_req(_obj_path, cast_obj);
3381
3382 // A merge of null or Casted-NotNull obj
3383 Node* res = _gvn.transform(phi);
3384
3385 // Note I do NOT always 'replace_in_map(obj,result)' here.
3386 // if( tk->klass()->can_be_primary_super() )
3387 // This means that if I successfully store an Object into an array-of-String
3388 // I 'forget' that the Object is really now known to be a String. I have to
3389 // do this because we don't have true union types for interfaces - if I store
3390 // a Baz into an array-of-Interface and then tell the optimizer it's an
3391 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3392 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3393 // replace_in_map( obj, res );
3394
3395 // Return final merged results
3396 set_control( _gvn.transform(region) );
3397 record_for_igvn(region);
3398
3399 return record_profiled_receiver_for_speculation(res);
3400 }
3401
3402 //------------------------------next_monitor-----------------------------------
3403 // What number should be given to the next monitor?
3404 int GraphKit::next_monitor() {
3405 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3406 int next = current + C->sync_stack_slots();
3407 // Keep the toplevel high water mark current:
3408 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3409 return current;
3410 }
3411
3412 //------------------------------insert_mem_bar---------------------------------
3413 // Memory barrier to avoid floating things around
3414 // The membar serves as a pinch point between both control and all memory slices.
3415 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3416 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3417 mb->init_req(TypeFunc::Control, control());
3418 mb->init_req(TypeFunc::Memory, reset_memory());
3419 Node* membar = _gvn.transform(mb);
3511 lock->create_lock_counter(map()->jvms());
3512 increment_counter(lock->counter()->addr());
3513 }
3514 #endif
3515
3516 return flock;
3517 }
3518
3519
3520 //------------------------------shared_unlock----------------------------------
3521 // Emit unlocking code.
3522 void GraphKit::shared_unlock(Node* box, Node* obj) {
3523 // bci is either a monitorenter bc or InvocationEntryBci
3524 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3525 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3526
3527 if (stopped()) { // Dead monitor?
3528 map()->pop_monitor(); // Kill monitor from debug info
3529 return;
3530 }
3531
3532 // Memory barrier to avoid floating things down past the locked region
3533 insert_mem_bar(Op_MemBarReleaseLock);
3534
3535 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3536 UnlockNode *unlock = new UnlockNode(C, tf);
3537 #ifdef ASSERT
3538 unlock->set_dbg_jvms(sync_jvms());
3539 #endif
3540 uint raw_idx = Compile::AliasIdxRaw;
3541 unlock->init_req( TypeFunc::Control, control() );
3542 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3543 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3544 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3545 unlock->init_req( TypeFunc::ReturnAdr, top() );
3546
3547 unlock->init_req(TypeFunc::Parms + 0, obj);
3548 unlock->init_req(TypeFunc::Parms + 1, box);
3549 unlock = _gvn.transform(unlock)->as_Unlock();
3550
3551 Node* mem = reset_memory();
3552
3553 // unlock has no side-effects, sets few values
3554 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3555
3556 // Kill monitor from debug info
3557 map()->pop_monitor( );
3558 }
3559
3560 //-------------------------------get_layout_helper-----------------------------
3561 // If the given klass is a constant or known to be an array,
3562 // fetch the constant layout helper value into constant_value
3563 // and return null. Otherwise, load the non-constant
3564 // layout helper value, and return the node which represents it.
3565 // This two-faced routine is useful because allocation sites
3566 // almost always feature constant types.
3567 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3568 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
3569 if (!StressReflectiveCode && klass_t != nullptr) {
3570 bool xklass = klass_t->klass_is_exact();
3571 if (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM)) {
3572 jint lhelper;
3573 if (klass_t->isa_aryklassptr()) {
3574 BasicType elem = klass_t->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
3575 if (is_reference_type(elem, true)) {
3576 elem = T_OBJECT;
3577 }
3578 lhelper = Klass::array_layout_helper(elem);
3579 } else {
3580 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
3581 }
3582 if (lhelper != Klass::_lh_neutral_value) {
3583 constant_value = lhelper;
3584 return (Node*) nullptr;
3585 }
3586 }
3587 }
3588 constant_value = Klass::_lh_neutral_value; // put in a known value
3589 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3590 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3591 }
3592
3593 // We just put in an allocate/initialize with a big raw-memory effect.
3594 // Hook selected additional alias categories on the initialization.
3595 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3596 MergeMemNode* init_in_merge,
3597 Node* init_out_raw) {
3598 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3599 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3600
3601 Node* prevmem = kit.memory(alias_idx);
3602 init_in_merge->set_memory_at(alias_idx, prevmem);
3603 kit.set_memory(init_out_raw, alias_idx);
3604 }
3605
3606 //---------------------------set_output_for_allocation-------------------------
3607 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3608 const TypeOopPtr* oop_type,
3609 bool deoptimize_on_exception) {
3610 int rawidx = Compile::AliasIdxRaw;
3611 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3612 add_safepoint_edges(alloc);
3613 Node* allocx = _gvn.transform(alloc);
3614 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3615 // create memory projection for i_o
3616 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3617 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3618
3619 // create a memory projection as for the normal control path
3620 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3621 set_memory(malloc, rawidx);
3622
3623 // a normal slow-call doesn't change i_o, but an allocation does
3624 // we create a separate i_o projection for the normal control path
3625 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3626 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3627
3628 // put in an initialization barrier
3629 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3630 rawoop)->as_Initialize();
3631 assert(alloc->initialization() == init, "2-way macro link must work");
3632 assert(init ->allocation() == alloc, "2-way macro link must work");
3633 {
3634 // Extract memory strands which may participate in the new object's
3635 // initialization, and source them from the new InitializeNode.
3636 // This will allow us to observe initializations when they occur,
3637 // and link them properly (as a group) to the InitializeNode.
3638 assert(init->in(InitializeNode::Memory) == malloc, "");
3639 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3640 init->set_req(InitializeNode::Memory, minit_in);
3641 record_for_igvn(minit_in); // fold it up later, if possible
3642 Node* minit_out = memory(rawidx);
3643 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3644 // Add an edge in the MergeMem for the header fields so an access
3645 // to one of those has correct memory state
3646 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
3647 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
3648 if (oop_type->isa_aryptr()) {
3649 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3650 int elemidx = C->get_alias_index(telemref);
3651 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3652 } else if (oop_type->isa_instptr()) {
3653 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
3654 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3655 ciField* field = ik->nonstatic_field_at(i);
3656 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
3657 continue; // do not bother to track really large numbers of fields
3658 // Find (or create) the alias category for this field:
3659 int fieldidx = C->alias_type(field)->index();
3660 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3661 }
3662 }
3663 }
3664
3665 // Cast raw oop to the real thing...
3666 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3667 javaoop = _gvn.transform(javaoop);
3668 C->set_recent_alloc(control(), javaoop);
3669 assert(just_allocated_object(control()) == javaoop, "just allocated");
3670
3671 #ifdef ASSERT
3672 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3683 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3684 }
3685 }
3686 #endif //ASSERT
3687
3688 return javaoop;
3689 }
3690
3691 //---------------------------new_instance--------------------------------------
3692 // This routine takes a klass_node which may be constant (for a static type)
3693 // or may be non-constant (for reflective code). It will work equally well
3694 // for either, and the graph will fold nicely if the optimizer later reduces
3695 // the type to a constant.
3696 // The optional arguments are for specialized use by intrinsics:
3697 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3698 // - If 'return_size_val', report the total object size to the caller.
3699 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3700 Node* GraphKit::new_instance(Node* klass_node,
3701 Node* extra_slow_test,
3702 Node* *return_size_val,
3703 bool deoptimize_on_exception) {
3704 // Compute size in doublewords
3705 // The size is always an integral number of doublewords, represented
3706 // as a positive bytewise size stored in the klass's layout_helper.
3707 // The layout_helper also encodes (in a low bit) the need for a slow path.
3708 jint layout_con = Klass::_lh_neutral_value;
3709 Node* layout_val = get_layout_helper(klass_node, layout_con);
3710 int layout_is_con = (layout_val == nullptr);
3711
3712 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
3713 // Generate the initial go-slow test. It's either ALWAYS (return a
3714 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
3715 // case) a computed value derived from the layout_helper.
3716 Node* initial_slow_test = nullptr;
3717 if (layout_is_con) {
3718 assert(!StressReflectiveCode, "stress mode does not use these paths");
3719 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3720 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3721 } else { // reflective case
3722 // This reflective path is used by Unsafe.allocateInstance.
3723 // (It may be stress-tested by specifying StressReflectiveCode.)
3724 // Basically, we want to get into the VM is there's an illegal argument.
3725 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3726 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3727 if (extra_slow_test != intcon(0)) {
3728 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3729 }
3730 // (Macro-expander will further convert this to a Bool, if necessary.)
3741
3742 // Clear the low bits to extract layout_helper_size_in_bytes:
3743 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3744 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3745 size = _gvn.transform( new AndXNode(size, mask) );
3746 }
3747 if (return_size_val != nullptr) {
3748 (*return_size_val) = size;
3749 }
3750
3751 // This is a precise notnull oop of the klass.
3752 // (Actually, it need not be precise if this is a reflective allocation.)
3753 // It's what we cast the result to.
3754 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3755 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
3756 const TypeOopPtr* oop_type = tklass->as_instance_type();
3757
3758 // Now generate allocation code
3759
3760 // The entire memory state is needed for slow path of the allocation
3761 // since GC and deoptimization can happened.
3762 Node *mem = reset_memory();
3763 set_all_memory(mem); // Create new memory state
3764
3765 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3766 control(), mem, i_o(),
3767 size, klass_node,
3768 initial_slow_test);
3769
3770 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3771 }
3772
3773 //-------------------------------new_array-------------------------------------
3774 // helper for both newarray and anewarray
3775 // The 'length' parameter is (obviously) the length of the array.
3776 // The optional arguments are for specialized use by intrinsics:
3777 // - If 'return_size_val', report the non-padded array size (sum of header size
3778 // and array body) to the caller.
3779 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3780 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3781 Node* length, // number of array elements
3782 int nargs, // number of arguments to push back for uncommon trap
3783 Node* *return_size_val,
3784 bool deoptimize_on_exception) {
3785 jint layout_con = Klass::_lh_neutral_value;
3786 Node* layout_val = get_layout_helper(klass_node, layout_con);
3787 int layout_is_con = (layout_val == nullptr);
3788
3789 if (!layout_is_con && !StressReflectiveCode &&
3790 !too_many_traps(Deoptimization::Reason_class_check)) {
3791 // This is a reflective array creation site.
3792 // Optimistically assume that it is a subtype of Object[],
3793 // so that we can fold up all the address arithmetic.
3794 layout_con = Klass::array_layout_helper(T_OBJECT);
3795 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3796 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3797 { BuildCutout unless(this, bol_lh, PROB_MAX);
3798 inc_sp(nargs);
3799 uncommon_trap(Deoptimization::Reason_class_check,
3800 Deoptimization::Action_maybe_recompile);
3801 }
3802 layout_val = nullptr;
3803 layout_is_con = true;
3804 }
3805
3806 // Generate the initial go-slow test. Make sure we do not overflow
3807 // if length is huge (near 2Gig) or negative! We do not need
3808 // exact double-words here, just a close approximation of needed
3809 // double-words. We can't add any offset or rounding bits, lest we
3810 // take a size -1 of bytes and make it positive. Use an unsigned
3811 // compare, so negative sizes look hugely positive.
3812 int fast_size_limit = FastAllocateSizeLimit;
3813 if (layout_is_con) {
3814 assert(!StressReflectiveCode, "stress mode does not use these paths");
3815 // Increase the size limit if we have exact knowledge of array type.
3816 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3817 assert(fast_size_limit == 0 || count_leading_zeros(fast_size_limit) > static_cast<unsigned>(LogBytesPerLong - log2_esize),
3818 "fast_size_limit (%d) overflow when shifted left by %d", fast_size_limit, LogBytesPerLong - log2_esize);
3819 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3820 }
3821
3822 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3823 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3824
3825 // --- Size Computation ---
3826 // array_size = round_to_heap(array_header + (length << elem_shift));
3827 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3828 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3829 // The rounding mask is strength-reduced, if possible.
3830 int round_mask = MinObjAlignmentInBytes - 1;
3831 Node* header_size = nullptr;
3832 // (T_BYTE has the weakest alignment and size restrictions...)
3833 if (layout_is_con) {
3834 int hsize = Klass::layout_helper_header_size(layout_con);
3835 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3836 if ((round_mask & ~right_n_bits(eshift)) == 0)
3837 round_mask = 0; // strength-reduce it if it goes away completely
3838 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3839 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3840 assert(header_size_min <= hsize, "generic minimum is smallest");
3841 header_size = intcon(hsize);
3842 } else {
3843 Node* hss = intcon(Klass::_lh_header_size_shift);
3844 Node* hsm = intcon(Klass::_lh_header_size_mask);
3845 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
3846 header_size = _gvn.transform(new AndINode(header_size, hsm));
3847 }
3848
3849 Node* elem_shift = nullptr;
3850 if (layout_is_con) {
3851 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3852 if (eshift != 0)
3853 elem_shift = intcon(eshift);
3854 } else {
3855 // There is no need to mask or shift this value.
3856 // The semantics of LShiftINode include an implicit mask to 0x1F.
3857 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3858 elem_shift = layout_val;
3907 }
3908 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
3909
3910 if (return_size_val != nullptr) {
3911 // This is the size
3912 (*return_size_val) = non_rounded_size;
3913 }
3914
3915 Node* size = non_rounded_size;
3916 if (round_mask != 0) {
3917 Node* mask1 = MakeConX(round_mask);
3918 size = _gvn.transform(new AddXNode(size, mask1));
3919 Node* mask2 = MakeConX(~round_mask);
3920 size = _gvn.transform(new AndXNode(size, mask2));
3921 }
3922 // else if round_mask == 0, the size computation is self-rounding
3923
3924 // Now generate allocation code
3925
3926 // The entire memory state is needed for slow path of the allocation
3927 // since GC and deoptimization can happened.
3928 Node *mem = reset_memory();
3929 set_all_memory(mem); // Create new memory state
3930
3931 if (initial_slow_test->is_Bool()) {
3932 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3933 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3934 }
3935
3936 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3937 Node* valid_length_test = _gvn.intcon(1);
3938 if (ary_type->isa_aryptr()) {
3939 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
3940 jint max = TypeAryPtr::max_array_length(bt);
3941 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
3942 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
3943 }
3944
3945 // Create the AllocateArrayNode and its result projections
3946 AllocateArrayNode* alloc
3947 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3948 control(), mem, i_o(),
3949 size, klass_node,
3950 initial_slow_test,
3951 length, valid_length_test);
3952
3953 // Cast to correct type. Note that the klass_node may be constant or not,
3954 // and in the latter case the actual array type will be inexact also.
3955 // (This happens via a non-constant argument to inline_native_newArray.)
3956 // In any case, the value of klass_node provides the desired array type.
3957 const TypeInt* length_type = _gvn.find_int_type(length);
3958 if (ary_type->isa_aryptr() && length_type != nullptr) {
3959 // Try to get a better type than POS for the size
3960 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3961 }
3962
3963 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3964
3965 array_ideal_length(alloc, ary_type, true);
3966 return javaoop;
3967 }
3968
3969 // The following "Ideal_foo" functions are placed here because they recognize
3970 // the graph shapes created by the functions immediately above.
3971
3972 //---------------------------Ideal_allocation----------------------------------
4067 void GraphKit::add_parse_predicates(int nargs) {
4068 if (ShortRunningLongLoop) {
4069 // Will narrow the limit down with a cast node. Predicates added later may depend on the cast so should be last when
4070 // walking up from the loop.
4071 add_parse_predicate(Deoptimization::Reason_short_running_long_loop, nargs);
4072 }
4073 if (UseLoopPredicate) {
4074 add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4075 if (UseProfiledLoopPredicate) {
4076 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4077 }
4078 }
4079 if (UseAutoVectorizationPredicate) {
4080 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, nargs);
4081 }
4082 // Loop Limit Check Predicate should be near the loop.
4083 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4084 }
4085
4086 void GraphKit::sync_kit(IdealKit& ideal) {
4087 set_all_memory(ideal.merged_memory());
4088 set_i_o(ideal.i_o());
4089 set_control(ideal.ctrl());
4090 }
4091
4092 void GraphKit::final_sync(IdealKit& ideal) {
4093 // Final sync IdealKit and graphKit.
4094 sync_kit(ideal);
4095 }
4096
4097 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4098 Node* len = load_array_length(load_String_value(str, set_ctrl));
4099 Node* coder = load_String_coder(str, set_ctrl);
4100 // Divide length by 2 if coder is UTF16
4101 return _gvn.transform(new RShiftINode(len, coder));
4102 }
4103
4104 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4105 int value_offset = java_lang_String::value_offset();
4106 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4107 false, nullptr, 0);
4108 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4109 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4110 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4111 ciTypeArrayKlass::make(T_BYTE), true, 0);
4112 Node* p = basic_plus_adr(str, str, value_offset);
4113 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4114 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4115 return load;
4116 }
4117
4118 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4119 if (!CompactStrings) {
4120 return intcon(java_lang_String::CODER_UTF16);
4121 }
4122 int coder_offset = java_lang_String::coder_offset();
4123 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4124 false, nullptr, 0);
4125 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4126
4127 Node* p = basic_plus_adr(str, str, coder_offset);
4128 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4129 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4130 return load;
4131 }
4132
4133 void GraphKit::store_String_value(Node* str, Node* value) {
4134 int value_offset = java_lang_String::value_offset();
4135 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4136 false, nullptr, 0);
4137 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4138
4139 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4140 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4141 }
4142
4143 void GraphKit::store_String_coder(Node* str, Node* value) {
4144 int coder_offset = java_lang_String::coder_offset();
4145 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4146 false, nullptr, 0);
4147 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4148
4149 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4150 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4151 }
4152
4153 // Capture src and dst memory state with a MergeMemNode
4154 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4155 if (src_type == dst_type) {
4156 // Types are equal, we don't need a MergeMemNode
4157 return memory(src_type);
4158 }
4159 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4160 record_for_igvn(merge); // fold it up later, if possible
4161 int src_idx = C->get_alias_index(src_type);
4162 int dst_idx = C->get_alias_index(dst_type);
4163 merge->set_memory_at(src_idx, memory(src_idx));
4164 merge->set_memory_at(dst_idx, memory(dst_idx));
4165 return merge;
4166 }
4239 i_char->init_req(2, AddI(i_char, intcon(2)));
4240
4241 set_control(IfFalse(iff));
4242 set_memory(st, TypeAryPtr::BYTES);
4243 }
4244
4245 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4246 if (!field->is_constant()) {
4247 return nullptr; // Field not marked as constant.
4248 }
4249 ciInstance* holder = nullptr;
4250 if (!field->is_static()) {
4251 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4252 if (const_oop != nullptr && const_oop->is_instance()) {
4253 holder = const_oop->as_instance();
4254 }
4255 }
4256 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4257 /*is_unsigned_load=*/false);
4258 if (con_type != nullptr) {
4259 return makecon(con_type);
4260 }
4261 return nullptr;
4262 }
4263
4264 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4265 const TypeOopPtr* obj_type = obj->bottom_type()->isa_oopptr();
4266 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4267 if (obj_type != nullptr && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4268 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4269 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4270 return casted_obj;
4271 }
4272 return obj;
4273 }
|
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/register.hpp"
26 #include "ci/ciFlatArrayKlass.hpp"
27 #include "ci/ciInlineKlass.hpp"
28 #include "ci/ciObjArray.hpp"
29 #include "ci/ciUtilities.hpp"
30 #include "classfile/javaClasses.hpp"
31 #include "compiler/compileLog.hpp"
32 #include "gc/shared/barrierSet.hpp"
33 #include "gc/shared/c2/barrierSetC2.hpp"
34 #include "interpreter/interpreter.hpp"
35 #include "memory/resourceArea.hpp"
36 #include "oops/flatArrayKlass.hpp"
37 #include "opto/addnode.hpp"
38 #include "opto/castnode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/graphKit.hpp"
41 #include "opto/idealKit.hpp"
42 #include "opto/inlinetypenode.hpp"
43 #include "opto/intrinsicnode.hpp"
44 #include "opto/locknode.hpp"
45 #include "opto/machnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/opaquenode.hpp"
48 #include "opto/parse.hpp"
49 #include "opto/rootnode.hpp"
50 #include "opto/runtime.hpp"
51 #include "opto/subtypenode.hpp"
52 #include "runtime/deoptimization.hpp"
53 #include "runtime/sharedRuntime.hpp"
54 #include "runtime/stubRoutines.hpp"
55 #include "utilities/bitMap.inline.hpp"
56 #include "utilities/growableArray.hpp"
57 #include "utilities/powerOfTwo.hpp"
58
59 //----------------------------GraphKit-----------------------------------------
60 // Main utility constructor.
61 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
62 : Phase(Phase::Parser),
63 _env(C->env()),
64 _gvn((gvn != nullptr) ? *gvn : *C->initial_gvn()),
65 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
66 {
67 assert(gvn == nullptr || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
68 _exceptions = jvms->map()->next_exception();
69 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
70 set_jvms(jvms);
71 #ifdef ASSERT
72 if (_gvn.is_IterGVN() != nullptr) {
73 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
74 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
75 _worklist_size = _gvn.C->igvn_worklist()->size();
76 }
77 #endif
78 }
79
80 // Private constructor for parser.
81 GraphKit::GraphKit()
82 : Phase(Phase::Parser),
83 _env(C->env()),
84 _gvn(*C->initial_gvn()),
85 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
86 {
87 _exceptions = nullptr;
88 set_map(nullptr);
89 DEBUG_ONLY(_sp = -99);
90 DEBUG_ONLY(set_bci(-99));
91 }
92
93
94
95 //---------------------------clean_stack---------------------------------------
96 // Clear away rubbish from the stack area of the JVM state.
97 // This destroys any arguments that may be waiting on the stack.
342 }
343 static inline void add_one_req(Node* dstphi, Node* src) {
344 assert(is_hidden_merge(dstphi), "must be a special merge node");
345 assert(!is_hidden_merge(src), "must not be a special merge node");
346 dstphi->add_req(src);
347 }
348
349 //-----------------------combine_exception_states------------------------------
350 // This helper function combines exception states by building phis on a
351 // specially marked state-merging region. These regions and phis are
352 // untransformed, and can build up gradually. The region is marked by
353 // having a control input of its exception map, rather than null. Such
354 // regions do not appear except in this function, and in use_exception_state.
355 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
356 if (failing_internal()) {
357 return; // dying anyway...
358 }
359 JVMState* ex_jvms = ex_map->_jvms;
360 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
361 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
362 // TODO 8325632 Re-enable
363 // assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
364 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
365 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
366 assert(ex_map->req() == phi_map->req(), "matching maps");
367 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
368 Node* hidden_merge_mark = root();
369 Node* region = phi_map->control();
370 MergeMemNode* phi_mem = phi_map->merged_memory();
371 MergeMemNode* ex_mem = ex_map->merged_memory();
372 if (region->in(0) != hidden_merge_mark) {
373 // The control input is not (yet) a specially-marked region in phi_map.
374 // Make it so, and build some phis.
375 region = new RegionNode(2);
376 _gvn.set_type(region, Type::CONTROL);
377 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
378 region->init_req(1, phi_map->control());
379 phi_map->set_control(region);
380 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
381 record_for_igvn(io_phi);
382 _gvn.set_type(io_phi, Type::ABIO);
383 phi_map->set_i_o(io_phi);
871 if (PrintMiscellaneous && (Verbose || WizardMode)) {
872 tty->print_cr("Zombie local %d: ", local);
873 jvms->dump();
874 }
875 return false;
876 }
877 }
878 }
879 return true;
880 }
881
882 #endif //ASSERT
883
884 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
885 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
886 ciMethod* cur_method = jvms->method();
887 int cur_bci = jvms->bci();
888 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
889 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
890 return Interpreter::bytecode_should_reexecute(code) ||
891 (is_anewarray && (code == Bytecodes::_multianewarray));
892 // Reexecute _multianewarray bytecode which was replaced with
893 // sequence of [a]newarray. See Parse::do_multianewarray().
894 //
895 // Note: interpreter should not have it set since this optimization
896 // is limited by dimensions and guarded by flag so in some cases
897 // multianewarray() runtime calls will be generated and
898 // the bytecode should not be reexecutes (stack will not be reset).
899 } else {
900 return false;
901 }
902 }
903
904 // Helper function for adding JVMState and debug information to node
905 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
906 // Add the safepoint edges to the call (or other safepoint).
907
908 // Make sure dead locals are set to top. This
909 // should help register allocation time and cut down on the size
910 // of the deoptimization information.
911 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
962 }
963
964 // Presize the call:
965 DEBUG_ONLY(uint non_debug_edges = call->req());
966 call->add_req_batch(top(), youngest_jvms->debug_depth());
967 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
968
969 // Set up edges so that the call looks like this:
970 // Call [state:] ctl io mem fptr retadr
971 // [parms:] parm0 ... parmN
972 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
973 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
974 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
975 // Note that caller debug info precedes callee debug info.
976
977 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
978 uint debug_ptr = call->req();
979
980 // Loop over the map input edges associated with jvms, add them
981 // to the call node, & reset all offsets to match call node array.
982
983 JVMState* callee_jvms = nullptr;
984 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
985 uint debug_end = debug_ptr;
986 uint debug_start = debug_ptr - in_jvms->debug_size();
987 debug_ptr = debug_start; // back up the ptr
988
989 uint p = debug_start; // walks forward in [debug_start, debug_end)
990 uint j, k, l;
991 SafePointNode* in_map = in_jvms->map();
992 out_jvms->set_map(call);
993
994 if (can_prune_locals) {
995 assert(in_jvms->method() == out_jvms->method(), "sanity");
996 // If the current throw can reach an exception handler in this JVMS,
997 // then we must keep everything live that can reach that handler.
998 // As a quick and dirty approximation, we look for any handlers at all.
999 if (in_jvms->method()->has_exception_handlers()) {
1000 can_prune_locals = false;
1001 }
1002 }
1003
1004 // Add the Locals
1005 k = in_jvms->locoff();
1006 l = in_jvms->loc_size();
1007 out_jvms->set_locoff(p);
1008 if (!can_prune_locals) {
1009 for (j = 0; j < l; j++) {
1010 call->set_req(p++, in_map->in(k + j));
1011 }
1012 } else {
1013 p += l; // already set to top above by add_req_batch
1014 }
1015
1016 // Add the Expression Stack
1017 k = in_jvms->stkoff();
1018 l = in_jvms->sp();
1019 out_jvms->set_stkoff(p);
1020 if (!can_prune_locals) {
1021 for (j = 0; j < l; j++) {
1022 call->set_req(p++, in_map->in(k + j));
1023 }
1024 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1025 // Divide stack into {S0,...,S1}, where S0 is set to top.
1026 uint s1 = stack_slots_not_pruned;
1027 stack_slots_not_pruned = 0; // for next iteration
1028 if (s1 > l) s1 = l;
1029 uint s0 = l - s1;
1030 p += s0; // skip the tops preinstalled by add_req_batch
1031 for (j = s0; j < l; j++)
1032 call->set_req(p++, in_map->in(k+j));
1033 } else {
1034 p += l; // already set to top above by add_req_batch
1035 }
1036
1037 // Add the Monitors
1038 k = in_jvms->monoff();
1039 l = in_jvms->mon_size();
1040 out_jvms->set_monoff(p);
1041 for (j = 0; j < l; j++)
1042 call->set_req(p++, in_map->in(k+j));
1043
1044 // Copy any scalar object fields.
1045 k = in_jvms->scloff();
1046 l = in_jvms->scl_size();
1047 out_jvms->set_scloff(p);
1048 for (j = 0; j < l; j++)
1049 call->set_req(p++, in_map->in(k+j));
1050
1051 // Finish the new jvms.
1052 out_jvms->set_endoff(p);
1053
1054 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1055 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1056 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1057 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1058 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1059 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1060
1061 // Update the two tail pointers in parallel.
1062 callee_jvms = out_jvms;
1063 out_jvms = out_jvms->caller();
1064 in_jvms = in_jvms->caller();
1065 }
1066
1067 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1068
1069 // Test the correctness of JVMState::debug_xxx accessors:
1070 assert(call->jvms()->debug_start() == non_debug_edges, "");
1071 assert(call->jvms()->debug_end() == call->req(), "");
1072 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1073 }
1074
1075 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1076 Bytecodes::Code code = java_bc();
1077 if (code == Bytecodes::_wide) {
1078 code = method()->java_code_at_bci(bci() + 1);
1079 }
1080
1081 if (code != Bytecodes::_illegal) {
1082 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1218 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1219 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1220 return _gvn.transform( new AndLNode(conv, mask) );
1221 }
1222
1223 Node* GraphKit::ConvL2I(Node* offset) {
1224 // short-circuit a common case
1225 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1226 if (offset_con != (jlong)Type::OffsetBot) {
1227 return intcon((int) offset_con);
1228 }
1229 return _gvn.transform( new ConvL2INode(offset));
1230 }
1231
1232 //-------------------------load_object_klass-----------------------------------
1233 Node* GraphKit::load_object_klass(Node* obj) {
1234 // Special-case a fresh allocation to avoid building nodes:
1235 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1236 if (akls != nullptr) return akls;
1237 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1238 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
1239 }
1240
1241 //-------------------------load_array_length-----------------------------------
1242 Node* GraphKit::load_array_length(Node* array) {
1243 // Special-case a fresh allocation to avoid building nodes:
1244 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1245 Node *alen;
1246 if (alloc == nullptr) {
1247 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1248 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1249 } else {
1250 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1251 }
1252 return alen;
1253 }
1254
1255 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1256 const TypeOopPtr* oop_type,
1257 bool replace_length_in_map) {
1258 Node* length = alloc->Ideal_length();
1267 replace_in_map(length, ccast);
1268 }
1269 return ccast;
1270 }
1271 }
1272 return length;
1273 }
1274
1275 //------------------------------do_null_check----------------------------------
1276 // Helper function to do a null pointer check. Returned value is
1277 // the incoming address with null casted away. You are allowed to use the
1278 // not-null value only if you are control dependent on the test.
1279 #ifndef PRODUCT
1280 extern uint explicit_null_checks_inserted,
1281 explicit_null_checks_elided;
1282 #endif
1283 Node* GraphKit::null_check_common(Node* value, BasicType type,
1284 // optional arguments for variations:
1285 bool assert_null,
1286 Node* *null_control,
1287 bool speculative,
1288 bool null_marker_check) {
1289 assert(!assert_null || null_control == nullptr, "not both at once");
1290 if (stopped()) return top();
1291 NOT_PRODUCT(explicit_null_checks_inserted++);
1292
1293 if (value->is_InlineType()) {
1294 // Null checking a scalarized but nullable inline type. Check the null marker
1295 // input instead of the oop input to avoid keeping buffer allocations alive.
1296 InlineTypeNode* vtptr = value->as_InlineType();
1297 while (vtptr->get_oop()->is_InlineType()) {
1298 vtptr = vtptr->get_oop()->as_InlineType();
1299 }
1300 null_check_common(vtptr->get_null_marker(), T_INT, assert_null, null_control, speculative, true);
1301 if (stopped()) {
1302 return top();
1303 }
1304 if (assert_null) {
1305 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts)
1306 // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass());
1307 // replace_in_map(value, vtptr);
1308 // return vtptr;
1309 replace_in_map(value, null());
1310 return null();
1311 }
1312 bool do_replace_in_map = (null_control == nullptr || (*null_control) == top());
1313 return cast_not_null(value, do_replace_in_map);
1314 }
1315
1316 // Construct null check
1317 Node *chk = nullptr;
1318 switch(type) {
1319 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1320 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1321 case T_ARRAY : // fall through
1322 type = T_OBJECT; // simplify further tests
1323 case T_OBJECT : {
1324 const Type *t = _gvn.type( value );
1325
1326 const TypeOopPtr* tp = t->isa_oopptr();
1327 if (tp != nullptr && !tp->is_loaded()
1328 // Only for do_null_check, not any of its siblings:
1329 && !assert_null && null_control == nullptr) {
1330 // Usually, any field access or invocation on an unloaded oop type
1331 // will simply fail to link, since the statically linked class is
1332 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1333 // the static class is loaded but the sharper oop type is not.
1334 // Rather than checking for this obscure case in lots of places,
1335 // we simply observe that a null check on an unloaded class
1399 }
1400 Node *oldcontrol = control();
1401 set_control(cfg);
1402 Node *res = cast_not_null(value);
1403 set_control(oldcontrol);
1404 NOT_PRODUCT(explicit_null_checks_elided++);
1405 return res;
1406 }
1407 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1408 if (cfg == nullptr) break; // Quit at region nodes
1409 depth++;
1410 }
1411 }
1412
1413 //-----------
1414 // Branch to failure if null
1415 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1416 Deoptimization::DeoptReason reason;
1417 if (assert_null) {
1418 reason = Deoptimization::reason_null_assert(speculative);
1419 } else if (type == T_OBJECT || null_marker_check) {
1420 reason = Deoptimization::reason_null_check(speculative);
1421 } else {
1422 reason = Deoptimization::Reason_div0_check;
1423 }
1424 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1425 // ciMethodData::has_trap_at will return a conservative -1 if any
1426 // must-be-null assertion has failed. This could cause performance
1427 // problems for a method after its first do_null_assert failure.
1428 // Consider using 'Reason_class_check' instead?
1429
1430 // To cause an implicit null check, we set the not-null probability
1431 // to the maximum (PROB_MAX). For an explicit check the probability
1432 // is set to a smaller value.
1433 if (null_control != nullptr || too_many_traps(reason)) {
1434 // probability is less likely
1435 ok_prob = PROB_LIKELY_MAG(3);
1436 } else if (!assert_null &&
1437 (ImplicitNullCheckThreshold > 0) &&
1438 method() != nullptr &&
1439 (method()->method_data()->trap_count(reason)
1473 }
1474
1475 if (assert_null) {
1476 // Cast obj to null on this path.
1477 replace_in_map(value, zerocon(type));
1478 return zerocon(type);
1479 }
1480
1481 // Cast obj to not-null on this path, if there is no null_control.
1482 // (If there is a null_control, a non-null value may come back to haunt us.)
1483 if (type == T_OBJECT) {
1484 Node* cast = cast_not_null(value, false);
1485 if (null_control == nullptr || (*null_control) == top())
1486 replace_in_map(value, cast);
1487 value = cast;
1488 }
1489
1490 return value;
1491 }
1492
1493 //------------------------------cast_not_null----------------------------------
1494 // Cast obj to not-null on this path
1495 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1496 if (obj->is_InlineType()) {
1497 Node* vt = obj->isa_InlineType()->clone_if_required(&gvn(), map(), do_replace_in_map);
1498 vt->as_InlineType()->set_null_marker(_gvn);
1499 vt = _gvn.transform(vt);
1500 if (do_replace_in_map) {
1501 replace_in_map(obj, vt);
1502 }
1503 return vt;
1504 }
1505 const Type *t = _gvn.type(obj);
1506 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1507 // Object is already not-null?
1508 if( t == t_not_null ) return obj;
1509
1510 Node* cast = new CastPPNode(control(), obj,t_not_null);
1511 cast = _gvn.transform( cast );
1512
1513 // Scan for instances of 'obj' in the current JVM mapping.
1514 // These instances are known to be not-null after the test.
1515 if (do_replace_in_map)
1516 replace_in_map(obj, cast);
1517
1518 return cast; // Return casted value
1519 }
1520
1521 Node* GraphKit::cast_to_non_larval(Node* obj) {
1522 const Type* obj_type = gvn().type(obj);
1523 if (obj->is_InlineType() || !obj_type->is_inlinetypeptr()) {
1524 return obj;
1525 }
1526
1527 Node* new_obj = InlineTypeNode::make_from_oop(this, obj, obj_type->inline_klass());
1528 replace_in_map(obj, new_obj);
1529 return new_obj;
1530 }
1531
1532 // Sometimes in intrinsics, we implicitly know an object is not null
1533 // (there's no actual null check) so we can cast it to not null. In
1534 // the course of optimizations, the input to the cast can become null.
1535 // In that case that data path will die and we need the control path
1536 // to become dead as well to keep the graph consistent. So we have to
1537 // add a check for null for which one branch can't be taken. It uses
1538 // an OpaqueNotNull node that will cause the check to be removed after loop
1539 // opts so the test goes away and the compiled code doesn't execute a
1540 // useless check.
1541 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1542 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1543 return value;
1544 }
1545 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1546 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1547 Node* opaq = _gvn.transform(new OpaqueNotNullNode(C, tst));
1548 IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1549 _gvn.set_type(iff, iff->Value(&_gvn));
1550 if (!tst->is_Con()) {
1551 record_for_igvn(iff);
1624 // These are layered on top of the factory methods in LoadNode and StoreNode,
1625 // and integrate with the parser's memory state and _gvn engine.
1626 //
1627
1628 // factory methods in "int adr_idx"
1629 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1630 MemNode::MemOrd mo,
1631 LoadNode::ControlDependency control_dependency,
1632 bool require_atomic_access,
1633 bool unaligned,
1634 bool mismatched,
1635 bool unsafe,
1636 uint8_t barrier_data) {
1637 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1638 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1639 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1640 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1641 Node* mem = memory(adr_idx);
1642 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1643 ld = _gvn.transform(ld);
1644
1645 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1646 // Improve graph before escape analysis and boxing elimination.
1647 record_for_igvn(ld);
1648 if (ld->is_DecodeN()) {
1649 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1650 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1651 // a Phi). Recording such cases is still perfectly sound, but may be
1652 // unnecessary and result in some minor IGVN overhead.
1653 record_for_igvn(ld->in(1));
1654 }
1655 }
1656 return ld;
1657 }
1658
1659 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1660 MemNode::MemOrd mo,
1661 bool require_atomic_access,
1662 bool unaligned,
1663 bool mismatched,
1664 bool unsafe,
1678 if (unsafe) {
1679 st->as_Store()->set_unsafe_access();
1680 }
1681 st->as_Store()->set_barrier_data(barrier_data);
1682 st = _gvn.transform(st);
1683 set_memory(st, adr_idx);
1684 // Back-to-back stores can only remove intermediate store with DU info
1685 // so push on worklist for optimizer.
1686 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1687 record_for_igvn(st);
1688
1689 return st;
1690 }
1691
1692 Node* GraphKit::access_store_at(Node* obj,
1693 Node* adr,
1694 const TypePtr* adr_type,
1695 Node* val,
1696 const Type* val_type,
1697 BasicType bt,
1698 DecoratorSet decorators,
1699 bool safe_for_replace,
1700 const InlineTypeNode* vt) {
1701 // Transformation of a value which could be null pointer (CastPP #null)
1702 // could be delayed during Parse (for example, in adjust_map_after_if()).
1703 // Execute transformation here to avoid barrier generation in such case.
1704 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1705 val = _gvn.makecon(TypePtr::NULL_PTR);
1706 }
1707
1708 if (stopped()) {
1709 return top(); // Dead path ?
1710 }
1711
1712 assert(val != nullptr, "not dead path");
1713 if (val->is_InlineType()) {
1714 // Store to non-flat field. Buffer the inline type and make sure
1715 // the store is re-executed if the allocation triggers deoptimization.
1716 PreserveReexecuteState preexecs(this);
1717 jvms()->set_should_reexecute(true);
1718 val = val->as_InlineType()->buffer(this, safe_for_replace);
1719 }
1720
1721 C2AccessValuePtr addr(adr, adr_type);
1722 C2AccessValue value(val, val_type);
1723 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr, nullptr, vt);
1724 if (access.is_raw()) {
1725 return _barrier_set->BarrierSetC2::store_at(access, value);
1726 } else {
1727 return _barrier_set->store_at(access, value);
1728 }
1729 }
1730
1731 Node* GraphKit::access_load_at(Node* obj, // containing obj
1732 Node* adr, // actual address to store val at
1733 const TypePtr* adr_type,
1734 const Type* val_type,
1735 BasicType bt,
1736 DecoratorSet decorators,
1737 Node* ctl) {
1738 if (stopped()) {
1739 return top(); // Dead path ?
1740 }
1741
1742 C2AccessValuePtr addr(adr, adr_type);
1743 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1744 if (access.is_raw()) {
1745 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1746 } else {
1747 return _barrier_set->load_at(access, val_type);
1748 }
1749 }
1750
1751 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1752 const Type* val_type,
1753 BasicType bt,
1754 DecoratorSet decorators) {
1755 if (stopped()) {
1756 return top(); // Dead path ?
1757 }
1758
1759 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1760 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1761 if (access.is_raw()) {
1762 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1763 } else {
1828 Node* new_val,
1829 const Type* value_type,
1830 BasicType bt,
1831 DecoratorSet decorators) {
1832 C2AccessValuePtr addr(adr, adr_type);
1833 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1834 if (access.is_raw()) {
1835 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1836 } else {
1837 return _barrier_set->atomic_add_at(access, new_val, value_type);
1838 }
1839 }
1840
1841 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1842 return _barrier_set->clone(this, src, dst, size, is_array);
1843 }
1844
1845 //-------------------------array_element_address-------------------------
1846 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1847 const TypeInt* sizetype, Node* ctrl) {
1848 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
1849 uint shift;
1850 uint header;
1851 if (arytype->is_flat() && arytype->klass_is_exact()) {
1852 // We can only determine the flat array layout statically if the klass is exact. Otherwise, we could have different
1853 // value classes at runtime with a potentially different layout. The caller needs to fall back to call
1854 // load/store_unknown_inline_Type() at runtime. We could return a sentinel node for the non-exact case but that
1855 // might mess with other GVN transformations in between. Thus, we just continue in the else branch normally, even
1856 // though we don't need the address node in this case and throw it away again.
1857 shift = arytype->flat_log_elem_size();
1858 header = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT);
1859 } else {
1860 shift = exact_log2(type2aelembytes(elembt));
1861 header = arrayOopDesc::base_offset_in_bytes(elembt);
1862 }
1863
1864 // short-circuit a common case (saves lots of confusing waste motion)
1865 jint idx_con = find_int_con(idx, -1);
1866 if (idx_con >= 0) {
1867 intptr_t offset = header + ((intptr_t)idx_con << shift);
1868 return basic_plus_adr(ary, offset);
1869 }
1870
1871 // must be correct type for alignment purposes
1872 Node* base = basic_plus_adr(ary, header);
1873 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1874 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1875 return basic_plus_adr(ary, base, scale);
1876 }
1877
1878 Node* GraphKit::cast_to_flat_array(Node* array, ciInlineKlass* vk, bool is_null_free, bool is_not_null_free, bool is_atomic) {
1879 assert(vk->maybe_flat_in_array(), "element of type %s cannot be flat in array", vk->name()->as_utf8());
1880 if (!vk->has_nullable_atomic_layout()) {
1881 // Element does not have a nullable flat layout, cannot be nullable
1882 is_null_free = true;
1883 }
1884 if (!vk->has_atomic_layout() && !vk->has_non_atomic_layout()) {
1885 // Element does not have a null-free flat layout, cannot be null-free
1886 is_not_null_free = true;
1887 }
1888 if (is_null_free) {
1889 // TODO 8350865 Impossible type
1890 is_not_null_free = false;
1891 }
1892
1893 bool is_exact = is_null_free || is_not_null_free;
1894 ciArrayKlass* array_klass = ciArrayKlass::make(vk, is_null_free, is_atomic, true);
1895 assert(array_klass->is_elem_null_free() == is_null_free, "inconsistency");
1896 assert(array_klass->is_elem_atomic() == is_atomic, "inconsistency");
1897 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1898 arytype = arytype->cast_to_exactness(is_exact);
1899 arytype = arytype->cast_to_not_null_free(is_not_null_free);
1900 assert(arytype->is_null_free() == is_null_free, "inconsistency");
1901 assert(arytype->is_not_null_free() == is_not_null_free, "inconsistency");
1902 assert(arytype->is_atomic() == is_atomic, "inconsistency");
1903 return _gvn.transform(new CastPPNode(control(), array, arytype, ConstraintCastNode::StrongDependency));
1904 }
1905
1906 //-------------------------load_array_element-------------------------
1907 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1908 const Type* elemtype = arytype->elem();
1909 BasicType elembt = elemtype->array_element_basic_type();
1910 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1911 if (elembt == T_NARROWOOP) {
1912 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1913 }
1914 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1915 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1916 return ld;
1917 }
1918
1919 //-------------------------set_arguments_for_java_call-------------------------
1920 // Arguments (pre-popped from the stack) are taken from the JVMS.
1921 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) {
1922 PreserveReexecuteState preexecs(this);
1923 if (EnableValhalla) {
1924 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
1925 // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
1926 jvms()->set_should_reexecute(true);
1927 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
1928 inc_sp(arg_size);
1929 }
1930 // Add the call arguments
1931 const TypeTuple* domain = call->tf()->domain_sig();
1932 uint nargs = domain->cnt();
1933 int arg_num = 0;
1934 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1935 Node* arg = argument(i-TypeFunc::Parms);
1936 const Type* t = domain->field_at(i);
1937 // TODO 8284443 A static call to a mismatched method should still be scalarized
1938 if (t->is_inlinetypeptr() && !call->method()->get_Method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) {
1939 // We don't pass inline type arguments by reference but instead pass each field of the inline type
1940 if (!arg->is_InlineType()) {
1941 assert(_gvn.type(arg)->is_zero_type() && !t->inline_klass()->is_null_free(), "Unexpected argument type");
1942 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass());
1943 }
1944 InlineTypeNode* vt = arg->as_InlineType();
1945 vt->pass_fields(this, call, idx, true, !t->maybe_null());
1946 // If an inline type argument is passed as fields, attach the Method* to the call site
1947 // to be able to access the extended signature later via attached_method_before_pc().
1948 // For example, see CompiledMethod::preserve_callee_argument_oops().
1949 call->set_override_symbolic_info(true);
1950 // Register an calling convention dependency on the callee method to make sure that this method is deoptimized and
1951 // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched.
1952 C->dependencies()->assert_mismatch_calling_convention(call->method());
1953 arg_num++;
1954 continue;
1955 } else if (arg->is_InlineType()) {
1956 // Pass inline type argument via oop to callee
1957 arg = arg->as_InlineType()->buffer(this, true);
1958 }
1959 if (t != Type::HALF) {
1960 arg_num++;
1961 }
1962 call->init_req(idx++, arg);
1963 }
1964 }
1965
1966 //---------------------------set_edges_for_java_call---------------------------
1967 // Connect a newly created call into the current JVMS.
1968 // A return value node (if any) is returned from set_edges_for_java_call.
1969 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1970
1971 // Add the predefined inputs:
1972 call->init_req( TypeFunc::Control, control() );
1973 call->init_req( TypeFunc::I_O , i_o() );
1974 call->init_req( TypeFunc::Memory , reset_memory() );
1975 call->init_req( TypeFunc::FramePtr, frameptr() );
1976 call->init_req( TypeFunc::ReturnAdr, top() );
1977
1978 add_safepoint_edges(call, must_throw);
1979
1980 Node* xcall = _gvn.transform(call);
1981
1982 if (xcall == top()) {
1983 set_control(top());
1984 return;
1985 }
1986 assert(xcall == call, "call identity is stable");
1987
1988 // Re-use the current map to produce the result.
1989
1990 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1991 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1992 set_all_memory_call(xcall, separate_io_proj);
1993
1994 //return xcall; // no need, caller already has it
1995 }
1996
1997 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1998 if (stopped()) return top(); // maybe the call folded up?
1999
2000 // Note: Since any out-of-line call can produce an exception,
2001 // we always insert an I_O projection from the call into the result.
2002
2003 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
2004
2005 if (separate_io_proj) {
2006 // The caller requested separate projections be used by the fall
2007 // through and exceptional paths, so replace the projections for
2008 // the fall through path.
2009 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
2010 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
2011 }
2012
2013 // Capture the return value, if any.
2014 Node* ret;
2015 if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) {
2016 ret = top();
2017 } else if (call->tf()->returns_inline_type_as_fields()) {
2018 // Return of multiple values (inline type fields): we create a
2019 // InlineType node, each field is a projection from the call.
2020 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
2021 uint base_input = TypeFunc::Parms;
2022 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, false);
2023 } else {
2024 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
2025 ciType* t = call->method()->return_type();
2026 if (!t->is_loaded() && InlineTypeReturnedAsFields) {
2027 // The return type is unloaded but the callee might later be C2 compiled and then return
2028 // in scalarized form when the return type is loaded. Handle this similar to what we do in
2029 // PhaseMacroExpand::expand_mh_intrinsic_return by calling into the runtime to buffer.
2030 // It's a bit unfortunate because we will deopt anyway but the interpreter needs an oop.
2031 IdealKit ideal(this);
2032 IdealVariable res(ideal);
2033 ideal.declarations_done();
2034 ideal.if_then(ret, BoolTest::eq, ideal.makecon(TypePtr::NULL_PTR)); {
2035 // Return value is null
2036 ideal.set(res, ret);
2037 } ideal.else_(); {
2038 // Return value is non-null
2039 sync_kit(ideal);
2040
2041 // Change return type of call to scalarized return
2042 const TypeFunc* tf = call->_tf;
2043 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2044 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2045 call->_tf = new_tf;
2046 _gvn.set_type(call, call->Value(&_gvn));
2047 _gvn.set_type(ret, ret->Value(&_gvn));
2048
2049 Node* store_to_buf_call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
2050 OptoRuntime::store_inline_type_fields_Type(),
2051 StubRoutines::store_inline_type_fields_to_buf(),
2052 nullptr, TypePtr::BOTTOM, ret);
2053
2054 // We don't know how many values are returned. This assumes the
2055 // worst case, that all available registers are used.
2056 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2057 if (domain->field_at(i) == Type::HALF) {
2058 store_to_buf_call->init_req(i, top());
2059 continue;
2060 }
2061 Node* proj =_gvn.transform(new ProjNode(call, i));
2062 store_to_buf_call->init_req(i, proj);
2063 }
2064 make_slow_call_ex(store_to_buf_call, env()->Throwable_klass(), false);
2065
2066 Node* buf = _gvn.transform(new ProjNode(store_to_buf_call, TypeFunc::Parms));
2067 const Type* buf_type = TypeOopPtr::make_from_klass(t->as_klass())->join_speculative(TypePtr::NOTNULL);
2068 buf = _gvn.transform(new CheckCastPPNode(control(), buf, buf_type));
2069
2070 ideal.set(res, buf);
2071 ideal.sync_kit(this);
2072 } ideal.end_if();
2073 sync_kit(ideal);
2074 ret = _gvn.transform(ideal.value(res));
2075 }
2076 if (t->is_klass()) {
2077 const Type* type = TypeOopPtr::make_from_klass(t->as_klass());
2078 if (type->is_inlinetypeptr()) {
2079 ret = InlineTypeNode::make_from_oop(this, ret, type->inline_klass());
2080 }
2081 }
2082 }
2083
2084 return ret;
2085 }
2086
2087 //--------------------set_predefined_input_for_runtime_call--------------------
2088 // Reading and setting the memory state is way conservative here.
2089 // The real problem is that I am not doing real Type analysis on memory,
2090 // so I cannot distinguish card mark stores from other stores. Across a GC
2091 // point the Store Barrier and the card mark memory has to agree. I cannot
2092 // have a card mark store and its barrier split across the GC point from
2093 // either above or below. Here I get that to happen by reading ALL of memory.
2094 // A better answer would be to separate out card marks from other memory.
2095 // For now, return the input memory state, so that it can be reused
2096 // after the call, if this call has restricted memory effects.
2097 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
2098 // Set fixed predefined input arguments
2099 call->init_req(TypeFunc::Control, control());
2100 call->init_req(TypeFunc::I_O, top()); // does no i/o
2101 call->init_req(TypeFunc::ReturnAdr, top());
2102 if (call->is_CallLeafPure()) {
2103 call->init_req(TypeFunc::Memory, top());
2165 if (use->is_MergeMem()) {
2166 wl.push(use);
2167 }
2168 }
2169 }
2170
2171 // Replace the call with the current state of the kit.
2172 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
2173 JVMState* ejvms = nullptr;
2174 if (has_exceptions()) {
2175 ejvms = transfer_exceptions_into_jvms();
2176 }
2177
2178 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2179 ReplacedNodes replaced_nodes_exception;
2180 Node* ex_ctl = top();
2181
2182 SafePointNode* final_state = stop();
2183
2184 // Find all the needed outputs of this call
2185 CallProjections* callprojs = call->extract_projections(true, do_asserts);
2186
2187 Unique_Node_List wl;
2188 Node* init_mem = call->in(TypeFunc::Memory);
2189 Node* final_mem = final_state->in(TypeFunc::Memory);
2190 Node* final_ctl = final_state->in(TypeFunc::Control);
2191 Node* final_io = final_state->in(TypeFunc::I_O);
2192
2193 // Replace all the old call edges with the edges from the inlining result
2194 if (callprojs->fallthrough_catchproj != nullptr) {
2195 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2196 }
2197 if (callprojs->fallthrough_memproj != nullptr) {
2198 if (final_mem->is_MergeMem()) {
2199 // Parser's exits MergeMem was not transformed but may be optimized
2200 final_mem = _gvn.transform(final_mem);
2201 }
2202 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2203 add_mergemem_users_to_worklist(wl, final_mem);
2204 }
2205 if (callprojs->fallthrough_ioproj != nullptr) {
2206 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2207 }
2208
2209 // Replace the result with the new result if it exists and is used
2210 if (callprojs->resproj[0] != nullptr && result != nullptr) {
2211 // If the inlined code is dead, the result projections for an inline type returned as
2212 // fields have not been replaced. They will go away once the call is replaced by TOP below.
2213 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()),
2214 "unexpected number of results");
2215 C->gvn_replace_by(callprojs->resproj[0], result);
2216 }
2217
2218 if (ejvms == nullptr) {
2219 // No exception edges to simply kill off those paths
2220 if (callprojs->catchall_catchproj != nullptr) {
2221 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2222 }
2223 if (callprojs->catchall_memproj != nullptr) {
2224 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2225 }
2226 if (callprojs->catchall_ioproj != nullptr) {
2227 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2228 }
2229 // Replace the old exception object with top
2230 if (callprojs->exobj != nullptr) {
2231 C->gvn_replace_by(callprojs->exobj, C->top());
2232 }
2233 } else {
2234 GraphKit ekit(ejvms);
2235
2236 // Load my combined exception state into the kit, with all phis transformed:
2237 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2238 replaced_nodes_exception = ex_map->replaced_nodes();
2239
2240 Node* ex_oop = ekit.use_exception_state(ex_map);
2241
2242 if (callprojs->catchall_catchproj != nullptr) {
2243 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2244 ex_ctl = ekit.control();
2245 }
2246 if (callprojs->catchall_memproj != nullptr) {
2247 Node* ex_mem = ekit.reset_memory();
2248 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2249 add_mergemem_users_to_worklist(wl, ex_mem);
2250 }
2251 if (callprojs->catchall_ioproj != nullptr) {
2252 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2253 }
2254
2255 // Replace the old exception object with the newly created one
2256 if (callprojs->exobj != nullptr) {
2257 C->gvn_replace_by(callprojs->exobj, ex_oop);
2258 }
2259 }
2260
2261 // Disconnect the call from the graph
2262 call->disconnect_inputs(C);
2263 C->gvn_replace_by(call, C->top());
2264
2265 // Clean up any MergeMems that feed other MergeMems since the
2266 // optimizer doesn't like that.
2267 while (wl.size() > 0) {
2268 _gvn.transform(wl.pop());
2269 }
2270
2271 if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2272 replaced_nodes.apply(C, final_ctl);
2273 }
2274 if (!ex_ctl->is_top() && do_replaced_nodes) {
2275 replaced_nodes_exception.apply(C, ex_ctl);
2276 }
2277 }
2278
2279
2280 //------------------------------increment_counter------------------------------
2281 // for statistics: increment a VM counter by 1
2282
2283 void GraphKit::increment_counter(address counter_addr) {
2284 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2285 increment_counter(adr1);
2286 }
2287
2288 void GraphKit::increment_counter(Node* counter_addr) {
2289 Node* ctrl = control();
2290 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2291 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2451 *
2452 * @param n node that the type applies to
2453 * @param exact_kls type from profiling
2454 * @param maybe_null did profiling see null?
2455 *
2456 * @return node with improved type
2457 */
2458 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2459 const Type* current_type = _gvn.type(n);
2460 assert(UseTypeSpeculation, "type speculation must be on");
2461
2462 const TypePtr* speculative = current_type->speculative();
2463
2464 // Should the klass from the profile be recorded in the speculative type?
2465 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2466 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2467 const TypeOopPtr* xtype = tklass->as_instance_type();
2468 assert(xtype->klass_is_exact(), "Should be exact");
2469 // Any reason to believe n is not null (from this profiling or a previous one)?
2470 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2471 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2472 // record the new speculative type's depth
2473 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2474 speculative = speculative->with_inline_depth(jvms()->depth());
2475 } else if (current_type->would_improve_ptr(ptr_kind)) {
2476 // Profiling report that null was never seen so we can change the
2477 // speculative type to non null ptr.
2478 if (ptr_kind == ProfileAlwaysNull) {
2479 speculative = TypePtr::NULL_PTR;
2480 } else {
2481 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2482 const TypePtr* ptr = TypePtr::NOTNULL;
2483 if (speculative != nullptr) {
2484 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2485 } else {
2486 speculative = ptr;
2487 }
2488 }
2489 }
2490
2491 if (speculative != current_type->speculative()) {
2492 // Build a type with a speculative type (what we think we know
2493 // about the type but will need a guard when we use it)
2494 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2495 // We're changing the type, we need a new CheckCast node to carry
2496 // the new type. The new type depends on the control: what
2497 // profiling tells us is only valid from here as far as we can
2498 // tell.
2499 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2500 cast = _gvn.transform(cast);
2501 replace_in_map(n, cast);
2502 n = cast;
2503 }
2504
2505 return n;
2506 }
2507
2508 /**
2509 * Record profiling data from receiver profiling at an invoke with the
2510 * type system so that it can propagate it (speculation)
2511 *
2512 * @param n receiver node
2513 *
2514 * @return node with improved type
2515 */
2516 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2517 if (!UseTypeSpeculation) {
2518 return n;
2519 }
2520 ciKlass* exact_kls = profile_has_unique_klass();
2521 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2522 if ((java_bc() == Bytecodes::_checkcast ||
2523 java_bc() == Bytecodes::_instanceof ||
2524 java_bc() == Bytecodes::_aastore) &&
2525 method()->method_data()->is_mature()) {
2526 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2527 if (data != nullptr) {
2528 if (java_bc() == Bytecodes::_aastore) {
2529 ciKlass* array_type = nullptr;
2530 ciKlass* element_type = nullptr;
2531 ProfilePtrKind element_ptr = ProfileMaybeNull;
2532 bool flat_array = true;
2533 bool null_free_array = true;
2534 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2535 exact_kls = element_type;
2536 ptr_kind = element_ptr;
2537 } else {
2538 if (!data->as_BitData()->null_seen()) {
2539 ptr_kind = ProfileNeverNull;
2540 } else {
2541 if (TypeProfileCasts) {
2542 assert(data->is_ReceiverTypeData(), "bad profile data type");
2543 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2544 uint i = 0;
2545 for (; i < call->row_limit(); i++) {
2546 ciKlass* receiver = call->receiver(i);
2547 if (receiver != nullptr) {
2548 break;
2549 }
2550 }
2551 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2552 }
2553 }
2554 }
2555 }
2556 }
2557 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2558 }
2559
2560 /**
2561 * Record profiling data from argument profiling at an invoke with the
2562 * type system so that it can propagate it (speculation)
2563 *
2564 * @param dest_method target method for the call
2565 * @param bc what invoke bytecode is this?
2566 */
2567 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2568 if (!UseTypeSpeculation) {
2569 return;
2570 }
2571 const TypeFunc* tf = TypeFunc::make(dest_method);
2572 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2573 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2574 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2575 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2576 if (is_reference_type(targ->basic_type())) {
2577 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2578 ciKlass* better_type = nullptr;
2579 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2580 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2581 }
2582 i++;
2583 }
2584 }
2585 }
2586
2587 /**
2588 * Record profiling data from parameter profiling at an invoke with
2589 * the type system so that it can propagate it (speculation)
2590 */
2591 void GraphKit::record_profiled_parameters_for_speculation() {
2592 if (!UseTypeSpeculation) {
2593 return;
2594 }
2595 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2715 // The first null ends the list.
2716 Node* parm0, Node* parm1,
2717 Node* parm2, Node* parm3,
2718 Node* parm4, Node* parm5,
2719 Node* parm6, Node* parm7) {
2720 assert(call_addr != nullptr, "must not call null targets");
2721
2722 // Slow-path call
2723 bool is_leaf = !(flags & RC_NO_LEAF);
2724 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2725 if (call_name == nullptr) {
2726 assert(!is_leaf, "must supply name for leaf");
2727 call_name = OptoRuntime::stub_name(call_addr);
2728 }
2729 CallNode* call;
2730 if (!is_leaf) {
2731 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2732 } else if (flags & RC_NO_FP) {
2733 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2734 } else if (flags & RC_VECTOR){
2735 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2736 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2737 } else if (flags & RC_PURE) {
2738 call = new CallLeafPureNode(call_type, call_addr, call_name, adr_type);
2739 } else {
2740 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2741 }
2742
2743 // The following is similar to set_edges_for_java_call,
2744 // except that the memory effects of the call are restricted to AliasIdxRaw.
2745
2746 // Slow path call has no side-effects, uses few values
2747 bool wide_in = !(flags & RC_NARROW_MEM);
2748 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2749
2750 Node* prev_mem = nullptr;
2751 if (wide_in) {
2752 prev_mem = set_predefined_input_for_runtime_call(call);
2753 } else {
2754 assert(!wide_out, "narrow in => narrow out");
2755 Node* narrow_mem = memory(adr_type);
2756 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2757 }
2758
2759 // Hook each parm in order. Stop looking at the first null.
2760 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2761 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2762 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2763 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2764 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2765 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2766 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2767 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2768 /* close each nested if ===> */ } } } } } } } }
2769 assert(call->in(call->req()-1) != nullptr || (call->req()-1) > (TypeFunc::Parms+7), "must initialize all parms");
2770
2771 if (!is_leaf) {
2772 // Non-leaves can block and take safepoints:
2773 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2774 }
2775 // Non-leaves can throw exceptions:
2776 if (has_io) {
2777 call->set_req(TypeFunc::I_O, i_o());
2778 }
2779
2780 if (flags & RC_UNCOMMON) {
2781 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2782 // (An "if" probability corresponds roughly to an unconditional count.
2783 // Sort of.)
2784 call->set_cnt(PROB_UNLIKELY_MAG(4));
2785 }
2786
2787 Node* c = _gvn.transform(call);
2788 assert(c == call, "cannot disappear");
2789
2797
2798 if (has_io) {
2799 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2800 }
2801 return call;
2802
2803 }
2804
2805 // i2b
2806 Node* GraphKit::sign_extend_byte(Node* in) {
2807 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2808 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2809 }
2810
2811 // i2s
2812 Node* GraphKit::sign_extend_short(Node* in) {
2813 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2814 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2815 }
2816
2817
2818 //------------------------------merge_memory-----------------------------------
2819 // Merge memory from one path into the current memory state.
2820 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2821 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2822 Node* old_slice = mms.force_memory();
2823 Node* new_slice = mms.memory2();
2824 if (old_slice != new_slice) {
2825 PhiNode* phi;
2826 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2827 if (mms.is_empty()) {
2828 // clone base memory Phi's inputs for this memory slice
2829 assert(old_slice == mms.base_memory(), "sanity");
2830 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2831 _gvn.set_type(phi, Type::MEMORY);
2832 for (uint i = 1; i < phi->req(); i++) {
2833 phi->init_req(i, old_slice->in(i));
2834 }
2835 } else {
2836 phi = old_slice->as_Phi(); // Phi was generated already
2837 }
2894 gvn.transform(iff);
2895 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2896 return iff;
2897 }
2898
2899 //-------------------------------gen_subtype_check-----------------------------
2900 // Generate a subtyping check. Takes as input the subtype and supertype.
2901 // Returns 2 values: sets the default control() to the true path and returns
2902 // the false path. Only reads invariant memory; sets no (visible) memory.
2903 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2904 // but that's not exposed to the optimizer. This call also doesn't take in an
2905 // Object; if you wish to check an Object you need to load the Object's class
2906 // prior to coming here.
2907 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
2908 ciMethod* method, int bci) {
2909 Compile* C = gvn.C;
2910 if ((*ctrl)->is_top()) {
2911 return C->top();
2912 }
2913
2914 const TypeKlassPtr* klass_ptr_type = gvn.type(superklass)->is_klassptr();
2915 // For a direct pointer comparison, we need the refined array klass pointer
2916 Node* vm_superklass = superklass;
2917 if (klass_ptr_type->isa_aryklassptr() && klass_ptr_type->klass_is_exact()) {
2918 vm_superklass = gvn.makecon(klass_ptr_type->is_aryklassptr()->refined_array_klass_ptr());
2919 }
2920
2921 // Fast check for identical types, perhaps identical constants.
2922 // The types can even be identical non-constants, in cases
2923 // involving Array.newInstance, Object.clone, etc.
2924 if (subklass == superklass)
2925 return C->top(); // false path is dead; no test needed.
2926
2927 if (gvn.type(superklass)->singleton()) {
2928 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2929 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
2930
2931 // In the common case of an exact superklass, try to fold up the
2932 // test before generating code. You may ask, why not just generate
2933 // the code and then let it fold up? The answer is that the generated
2934 // code will necessarily include null checks, which do not always
2935 // completely fold away. If they are also needless, then they turn
2936 // into a performance loss. Example:
2937 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2938 // Here, the type of 'fa' is often exact, so the store check
2939 // of fa[1]=x will fold up, without testing the nullness of x.
2940 //
2941 // At macro expansion, we would have already folded the SubTypeCheckNode
2942 // being expanded here because we always perform the static sub type
2943 // check in SubTypeCheckNode::sub() regardless of whether
2944 // StressReflectiveCode is set or not. We can therefore skip this
2945 // static check when StressReflectiveCode is on.
2946 switch (C->static_subtype_check(superk, subk)) {
2947 case Compile::SSC_always_false:
2948 {
2949 Node* always_fail = *ctrl;
2950 *ctrl = gvn.C->top();
2951 return always_fail;
2952 }
2953 case Compile::SSC_always_true:
2954 return C->top();
2955 case Compile::SSC_easy_test:
2956 {
2957 // Just do a direct pointer compare and be done.
2958 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2959 *ctrl = gvn.transform(new IfTrueNode(iff));
2960 return gvn.transform(new IfFalseNode(iff));
2961 }
2962 case Compile::SSC_full_test:
2963 break;
2964 default:
2965 ShouldNotReachHere();
2966 }
2967 }
2968
2969 // %%% Possible further optimization: Even if the superklass is not exact,
2970 // if the subklass is the unique subtype of the superklass, the check
2971 // will always succeed. We could leave a dependency behind to ensure this.
2972
2973 // First load the super-klass's check-offset
2974 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2975 Node* m = C->immutable_memory();
2976 Node *chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2977 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2978 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
3016 gvn.record_for_igvn(r_ok_subtype);
3017
3018 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
3019 // SubTypeCheck node
3020 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
3021 ciCallProfile profile = method->call_profile_at_bci(bci);
3022 float total_prob = 0;
3023 for (int i = 0; profile.has_receiver(i); ++i) {
3024 float prob = profile.receiver_prob(i);
3025 total_prob += prob;
3026 }
3027 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
3028 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
3029 for (int i = 0; profile.has_receiver(i); ++i) {
3030 ciKlass* klass = profile.receiver(i);
3031 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
3032 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
3033 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
3034 continue;
3035 }
3036 if (klass_t->isa_aryklassptr()) {
3037 // For a direct pointer comparison, we need the refined array klass pointer
3038 klass_t = klass_t->is_aryklassptr()->refined_array_klass_ptr();
3039 }
3040 float prob = profile.receiver_prob(i);
3041 ConNode* klass_node = gvn.makecon(klass_t);
3042 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
3043 Node* iftrue = gvn.transform(new IfTrueNode(iff));
3044
3045 if (result == Compile::SSC_always_true) {
3046 r_ok_subtype->add_req(iftrue);
3047 } else {
3048 assert(result == Compile::SSC_always_false, "");
3049 r_not_subtype->add_req(iftrue);
3050 }
3051 *ctrl = gvn.transform(new IfFalseNode(iff));
3052 }
3053 }
3054 }
3055
3056 // See if we get an immediate positive hit. Happens roughly 83% of the
3057 // time. Test to see if the value loaded just previously from the subklass
3058 // is exactly the superklass.
3059 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
3073 igvn->remove_globally_dead_node(r_not_subtype);
3074 }
3075 return not_subtype_ctrl;
3076 }
3077
3078 r_ok_subtype->init_req(1, iftrue1);
3079
3080 // Check for immediate negative hit. Happens roughly 11% of the time (which
3081 // is roughly 63% of the remaining cases). Test to see if the loaded
3082 // check-offset points into the subklass display list or the 1-element
3083 // cache. If it points to the display (and NOT the cache) and the display
3084 // missed then it's not a subtype.
3085 Node *cacheoff = gvn.intcon(cacheoff_con);
3086 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
3087 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
3088 *ctrl = gvn.transform(new IfFalseNode(iff2));
3089
3090 // Check for self. Very rare to get here, but it is taken 1/3 the time.
3091 // No performance impact (too rare) but allows sharing of secondary arrays
3092 // which has some footprint reduction.
3093 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
3094 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
3095 *ctrl = gvn.transform(new IfFalseNode(iff3));
3096
3097 // -- Roads not taken here: --
3098 // We could also have chosen to perform the self-check at the beginning
3099 // of this code sequence, as the assembler does. This would not pay off
3100 // the same way, since the optimizer, unlike the assembler, can perform
3101 // static type analysis to fold away many successful self-checks.
3102 // Non-foldable self checks work better here in second position, because
3103 // the initial primary superclass check subsumes a self-check for most
3104 // types. An exception would be a secondary type like array-of-interface,
3105 // which does not appear in its own primary supertype display.
3106 // Finally, we could have chosen to move the self-check into the
3107 // PartialSubtypeCheckNode, and from there out-of-line in a platform
3108 // dependent manner. But it is worthwhile to have the check here,
3109 // where it can be perhaps be optimized. The cost in code space is
3110 // small (register compare, branch).
3111
3112 // Now do a linear scan of the secondary super-klass array. Again, no real
3113 // performance impact (too rare) but it's gotta be done.
3114 // Since the code is rarely used, there is no penalty for moving it
3115 // out of line, and it can only improve I-cache density.
3116 // The decision to inline or out-of-line this final check is platform
3117 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
3118 Node* psc = gvn.transform(
3119 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
3120
3121 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
3122 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
3123 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
3124
3125 // Return false path; set default control to true path.
3126 *ctrl = gvn.transform(r_ok_subtype);
3127 return gvn.transform(r_not_subtype);
3128 }
3129
3130 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
3131 const Type* sub_t = _gvn.type(obj_or_subklass);
3132 if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) {
3133 sub_t = TypeKlassPtr::make(sub_t->inline_klass());
3134 obj_or_subklass = makecon(sub_t);
3135 }
3136 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
3137 if (expand_subtype_check) {
3138 MergeMemNode* mem = merged_memory();
3139 Node* ctrl = control();
3140 Node* subklass = obj_or_subklass;
3141 if (!sub_t->isa_klassptr()) {
3142 subklass = load_object_klass(obj_or_subklass);
3143 }
3144
3145 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
3146 set_control(ctrl);
3147 return n;
3148 }
3149
3150 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
3151 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
3152 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3153 set_control(_gvn.transform(new IfTrueNode(iff)));
3154 return _gvn.transform(new IfFalseNode(iff));
3155 }
3156
3157 // Profile-driven exact type check:
3158 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3159 float prob, Node* *casted_receiver) {
3160 assert(!klass->is_interface(), "no exact type check on interfaces");
3161 Node* fail = top();
3162 const Type* rec_t = _gvn.type(receiver);
3163 if (rec_t->is_inlinetypeptr()) {
3164 if (klass->equals(rec_t->inline_klass())) {
3165 (*casted_receiver) = receiver; // Always passes
3166 } else {
3167 (*casted_receiver) = top(); // Always fails
3168 fail = control();
3169 set_control(top());
3170 }
3171 return fail;
3172 }
3173 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3174 if (tklass->isa_aryklassptr()) {
3175 // For a direct pointer comparison, we need the refined array klass pointer
3176 tklass = tklass->is_aryklassptr()->refined_array_klass_ptr();
3177 }
3178 Node* recv_klass = load_object_klass(receiver);
3179 fail = type_check(recv_klass, tklass, prob);
3180
3181 if (!stopped()) {
3182 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3183 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
3184 assert(recv_xtype->klass_is_exact(), "");
3185
3186 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
3187 // Subsume downstream occurrences of receiver with a cast to
3188 // recv_xtype, since now we know what the type will be.
3189 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
3190 Node* res = _gvn.transform(cast);
3191 if (recv_xtype->is_inlinetypeptr()) {
3192 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3193 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
3194 }
3195 (*casted_receiver) = res;
3196 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3197 // (User must make the replace_in_map call.)
3198 }
3199 }
3200
3201 return fail;
3202 }
3203
3204 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3205 float prob) {
3206 Node* want_klass = makecon(tklass);
3207 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3208 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3209 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3210 set_control(_gvn.transform(new IfTrueNode (iff)));
3211 Node* fail = _gvn.transform(new IfFalseNode(iff));
3212 return fail;
3213 }
3214
3215 //------------------------------subtype_check_receiver-------------------------
3216 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3217 Node** casted_receiver) {
3218 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3219 Node* want_klass = makecon(tklass);
3220
3221 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3222
3223 // Ignore interface type information until interface types are properly tracked.
3224 if (!stopped() && !klass->is_interface()) {
3225 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3226 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3227 if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3228 Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type));
3229 if (recv_type->is_inlinetypeptr()) {
3230 cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass());
3231 }
3232 (*casted_receiver) = cast;
3233 }
3234 }
3235
3236 return slow_ctl;
3237 }
3238
3239 //------------------------------seems_never_null-------------------------------
3240 // Use null_seen information if it is available from the profile.
3241 // If we see an unexpected null at a type check we record it and force a
3242 // recompile; the offending check will be recompiled to handle nulls.
3243 // If we see several offending BCIs, then all checks in the
3244 // method will be recompiled.
3245 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3246 speculating = !_gvn.type(obj)->speculative_maybe_null();
3247 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3248 if (UncommonNullCast // Cutout for this technique
3249 && obj != null() // And not the -Xcomp stupid case?
3250 && !too_many_traps(reason)
3251 ) {
3252 if (speculating) {
3321
3322 //------------------------maybe_cast_profiled_receiver-------------------------
3323 // If the profile has seen exactly one type, narrow to exactly that type.
3324 // Subsequent type checks will always fold up.
3325 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3326 const TypeKlassPtr* require_klass,
3327 ciKlass* spec_klass,
3328 bool safe_for_replace) {
3329 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3330
3331 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3332
3333 // Make sure we haven't already deoptimized from this tactic.
3334 if (too_many_traps_or_recompiles(reason))
3335 return nullptr;
3336
3337 // (No, this isn't a call, but it's enough like a virtual call
3338 // to use the same ciMethod accessor to get the profile info...)
3339 // If we have a speculative type use it instead of profiling (which
3340 // may not help us)
3341 ciKlass* exact_kls = spec_klass;
3342 if (exact_kls == nullptr) {
3343 if (java_bc() == Bytecodes::_aastore) {
3344 ciKlass* array_type = nullptr;
3345 ciKlass* element_type = nullptr;
3346 ProfilePtrKind element_ptr = ProfileMaybeNull;
3347 bool flat_array = true;
3348 bool null_free_array = true;
3349 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3350 exact_kls = element_type;
3351 } else {
3352 exact_kls = profile_has_unique_klass();
3353 }
3354 }
3355 if (exact_kls != nullptr) {// no cast failures here
3356 if (require_klass == nullptr ||
3357 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3358 // If we narrow the type to match what the type profile sees or
3359 // the speculative type, we can then remove the rest of the
3360 // cast.
3361 // This is a win, even if the exact_kls is very specific,
3362 // because downstream operations, such as method calls,
3363 // will often benefit from the sharper type.
3364 Node* exact_obj = not_null_obj; // will get updated in place...
3365 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3366 &exact_obj);
3367 { PreserveJVMState pjvms(this);
3368 set_control(slow_ctl);
3369 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3370 }
3371 if (safe_for_replace) {
3372 replace_in_map(not_null_obj, exact_obj);
3373 }
3374 return exact_obj;
3464 // If not_null_obj is dead, only null-path is taken
3465 if (stopped()) { // Doing instance-of on a null?
3466 set_control(null_ctl);
3467 return intcon(0);
3468 }
3469 region->init_req(_null_path, null_ctl);
3470 phi ->init_req(_null_path, intcon(0)); // Set null path value
3471 if (null_ctl == top()) {
3472 // Do this eagerly, so that pattern matches like is_diamond_phi
3473 // will work even during parsing.
3474 assert(_null_path == PATH_LIMIT-1, "delete last");
3475 region->del_req(_null_path);
3476 phi ->del_req(_null_path);
3477 }
3478
3479 // Do we know the type check always succeed?
3480 bool known_statically = false;
3481 if (_gvn.type(superklass)->singleton()) {
3482 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3483 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3484 if (subk != nullptr && subk->is_loaded()) {
3485 int static_res = C->static_subtype_check(superk, subk);
3486 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3487 }
3488 }
3489
3490 if (!known_statically) {
3491 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3492 // We may not have profiling here or it may not help us. If we
3493 // have a speculative type use it to perform an exact cast.
3494 ciKlass* spec_obj_type = obj_type->speculative_type();
3495 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3496 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3497 if (stopped()) { // Profile disagrees with this path.
3498 set_control(null_ctl); // Null is the only remaining possibility.
3499 return intcon(0);
3500 }
3501 if (cast_obj != nullptr) {
3502 not_null_obj = cast_obj;
3503 }
3504 }
3520 record_for_igvn(region);
3521
3522 // If we know the type check always succeeds then we don't use the
3523 // profiling data at this bytecode. Don't lose it, feed it to the
3524 // type system as a speculative type.
3525 if (safe_for_replace) {
3526 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3527 replace_in_map(obj, casted_obj);
3528 }
3529
3530 return _gvn.transform(phi);
3531 }
3532
3533 //-------------------------------gen_checkcast---------------------------------
3534 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3535 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3536 // uncommon-trap paths work. Adjust stack after this call.
3537 // If failure_control is supplied and not null, it is filled in with
3538 // the control edge for the cast failure. Otherwise, an appropriate
3539 // uncommon trap or exception is thrown.
3540 Node* GraphKit::gen_checkcast(Node* obj, Node* superklass, Node* *failure_control, bool null_free, bool maybe_larval) {
3541 kill_dead_locals(); // Benefit all the uncommon traps
3542 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3543 const Type* obj_type = _gvn.type(obj);
3544 if (obj_type->is_inlinetypeptr() && !obj_type->maybe_null() && klass_ptr_type->klass_is_exact() && obj_type->inline_klass() == klass_ptr_type->exact_klass(true)) {
3545 // Special case: larval inline objects must not be scalarized. They are also generally not
3546 // allowed to participate in most operations except as the first operand of putfield, or as an
3547 // argument to a constructor invocation with it being a receiver, Unsafe::putXXX with it being
3548 // the first argument, or Unsafe::finishPrivateBuffer. This allows us to aggressively scalarize
3549 // value objects in all other places. This special case comes from the limitation of the Java
3550 // language, Unsafe::makePrivateBuffer returns an Object that is checkcast-ed to the concrete
3551 // value type. We must do this first because C->static_subtype_check may do nothing when
3552 // StressReflectiveCode is set.
3553 return obj;
3554 }
3555
3556 // Else it must be a non-larval object
3557 obj = cast_to_non_larval(obj);
3558
3559 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3560 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3561 bool safe_for_replace = (failure_control == nullptr);
3562 assert(!null_free || toop->can_be_inline_type(), "must be an inline type pointer");
3563
3564 // Fast cutout: Check the case that the cast is vacuously true.
3565 // This detects the common cases where the test will short-circuit
3566 // away completely. We do this before we perform the null check,
3567 // because if the test is going to turn into zero code, we don't
3568 // want a residual null check left around. (Causes a slowdown,
3569 // for example, in some objArray manipulations, such as a[i]=a[j].)
3570 if (improved_klass_ptr_type->singleton()) {
3571 const TypeKlassPtr* kptr = nullptr;
3572 if (obj_type->isa_oop_ptr()) {
3573 kptr = obj_type->is_oopptr()->as_klass_type();
3574 } else if (obj->is_InlineType()) {
3575 ciInlineKlass* vk = obj_type->inline_klass();
3576 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0));
3577 }
3578
3579 if (kptr != nullptr) {
3580 switch (C->static_subtype_check(improved_klass_ptr_type, kptr)) {
3581 case Compile::SSC_always_true:
3582 // If we know the type check always succeed then we don't use
3583 // the profiling data at this bytecode. Don't lose it, feed it
3584 // to the type system as a speculative type.
3585 obj = record_profiled_receiver_for_speculation(obj);
3586 if (null_free) {
3587 assert(safe_for_replace, "must be");
3588 obj = null_check(obj);
3589 }
3590 assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineType(), "should have been scalarized");
3591 return obj;
3592 case Compile::SSC_always_false:
3593 if (null_free) {
3594 assert(safe_for_replace, "must be");
3595 obj = null_check(obj);
3596 }
3597 // It needs a null check because a null will *pass* the cast check.
3598 if (obj_type->isa_oopptr() != nullptr && !obj_type->is_oopptr()->maybe_null()) {
3599 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3600 Deoptimization::DeoptReason reason = is_aastore ?
3601 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3602 builtin_throw(reason);
3603 return top();
3604 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3605 return null_assert(obj);
3606 }
3607 break; // Fall through to full check
3608 default:
3609 break;
3610 }
3611 }
3612 }
3613
3614 ciProfileData* data = nullptr;
3615 if (failure_control == nullptr) { // use MDO in regular case only
3616 assert(java_bc() == Bytecodes::_aastore ||
3617 java_bc() == Bytecodes::_checkcast,
3618 "interpreter profiles type checks only for these BCs");
3619 if (method()->method_data()->is_mature()) {
3620 data = method()->method_data()->bci_to_data(bci());
3621 }
3622 }
3623
3624 // Make the merge point
3625 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3626 RegionNode* region = new RegionNode(PATH_LIMIT);
3627 Node* phi = new PhiNode(region, toop);
3628 _gvn.set_type(region, Type::CONTROL);
3629 _gvn.set_type(phi, toop);
3630
3631 C->set_has_split_ifs(true); // Has chance for split-if optimization
3632
3633 // Use null-cast information if it is available
3634 bool speculative_not_null = false;
3635 bool never_see_null = ((failure_control == nullptr) // regular case only
3636 && seems_never_null(obj, data, speculative_not_null));
3637
3638 if (obj->is_InlineType()) {
3639 // Re-execute if buffering during triggers deoptimization
3640 PreserveReexecuteState preexecs(this);
3641 jvms()->set_should_reexecute(true);
3642 obj = obj->as_InlineType()->buffer(this, safe_for_replace);
3643 }
3644
3645 // Null check; get casted pointer; set region slot 3
3646 Node* null_ctl = top();
3647 Node* not_null_obj = nullptr;
3648 if (null_free) {
3649 assert(safe_for_replace, "must be");
3650 not_null_obj = null_check(obj);
3651 } else {
3652 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3653 }
3654
3655 // If not_null_obj is dead, only null-path is taken
3656 if (stopped()) { // Doing instance-of on a null?
3657 set_control(null_ctl);
3658 if (toop->is_inlinetypeptr()) {
3659 return InlineTypeNode::make_null(_gvn, toop->inline_klass());
3660 }
3661 return null();
3662 }
3663 region->init_req(_null_path, null_ctl);
3664 phi ->init_req(_null_path, null()); // Set null path value
3665 if (null_ctl == top()) {
3666 // Do this eagerly, so that pattern matches like is_diamond_phi
3667 // will work even during parsing.
3668 assert(_null_path == PATH_LIMIT-1, "delete last");
3669 region->del_req(_null_path);
3670 phi ->del_req(_null_path);
3671 }
3672
3673 Node* cast_obj = nullptr;
3674 if (improved_klass_ptr_type->klass_is_exact()) {
3675 // The following optimization tries to statically cast the speculative type of the object
3676 // (for example obtained during profiling) to the type of the superklass and then do a
3677 // dynamic check that the type of the object is what we expect. To work correctly
3678 // for checkcast and aastore the type of superklass should be exact.
3679 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3680 // We may not have profiling here or it may not help us. If we have
3681 // a speculative type use it to perform an exact cast.
3682 ciKlass* spec_obj_type = obj_type->speculative_type();
3683 if (spec_obj_type != nullptr || data != nullptr) {
3684 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3685 if (cast_obj != nullptr) {
3686 if (failure_control != nullptr) // failure is now impossible
3687 (*failure_control) = top();
3688 // adjust the type of the phi to the exact klass:
3689 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3690 }
3691 }
3692 }
3693
3694 if (cast_obj == nullptr) {
3695 // Generate the subtype check
3696 Node* improved_superklass = superklass;
3697 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3698 // Only improve the super class for constants which allows subsequent sub type checks to possibly be commoned up.
3699 // The other non-constant cases cannot be improved with a cast node here since they could be folded to top.
3700 // Additionally, the benefit would only be minor in non-constant cases.
3701 improved_superklass = makecon(improved_klass_ptr_type);
3702 }
3703 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3704 // Plug in success path into the merge
3705 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3706 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3707 if (failure_control == nullptr) {
3708 if (not_subtype_ctrl != top()) { // If failure is possible
3709 PreserveJVMState pjvms(this);
3710 set_control(not_subtype_ctrl);
3711 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3712 Deoptimization::DeoptReason reason = is_aastore ?
3713 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3714 builtin_throw(reason);
3715 }
3716 } else {
3717 (*failure_control) = not_subtype_ctrl;
3718 }
3719 }
3720
3721 region->init_req(_obj_path, control());
3722 phi ->init_req(_obj_path, cast_obj);
3723
3724 // A merge of null or Casted-NotNull obj
3725 Node* res = _gvn.transform(phi);
3726
3727 // Note I do NOT always 'replace_in_map(obj,result)' here.
3728 // if( tk->klass()->can_be_primary_super() )
3729 // This means that if I successfully store an Object into an array-of-String
3730 // I 'forget' that the Object is really now known to be a String. I have to
3731 // do this because we don't have true union types for interfaces - if I store
3732 // a Baz into an array-of-Interface and then tell the optimizer it's an
3733 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3734 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3735 // replace_in_map( obj, res );
3736
3737 // Return final merged results
3738 set_control( _gvn.transform(region) );
3739 record_for_igvn(region);
3740
3741 bool not_inline = !toop->can_be_inline_type();
3742 bool not_flat_in_array = !UseArrayFlattening || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->maybe_flat_in_array());
3743 if (EnableValhalla && (not_inline || not_flat_in_array)) {
3744 // Check if obj has been loaded from an array
3745 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3746 Node* array = nullptr;
3747 if (obj->isa_Load()) {
3748 Node* address = obj->in(MemNode::Address);
3749 if (address->isa_AddP()) {
3750 array = address->as_AddP()->in(AddPNode::Base);
3751 }
3752 } else if (obj->is_Phi()) {
3753 Node* region = obj->in(0);
3754 // TODO make this more robust (see JDK-8231346)
3755 if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) {
3756 IfNode* iff = region->in(2)->in(0)->isa_If();
3757 if (iff != nullptr) {
3758 iff->is_flat_array_check(&_gvn, &array);
3759 }
3760 }
3761 }
3762 if (array != nullptr) {
3763 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3764 if (ary_t != nullptr) {
3765 if (!ary_t->is_not_null_free() && !ary_t->is_null_free() && not_inline) {
3766 // Casting array element to a non-inline-type, mark array as not null-free.
3767 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3768 replace_in_map(array, cast);
3769 array = cast;
3770 }
3771 if (!ary_t->is_not_flat() && !ary_t->is_flat() && not_flat_in_array) {
3772 // Casting array element to a non-flat-in-array type, mark array as not flat.
3773 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3774 replace_in_map(array, cast);
3775 array = cast;
3776 }
3777 }
3778 }
3779 }
3780
3781 if (!stopped() && !res->is_InlineType()) {
3782 res = record_profiled_receiver_for_speculation(res);
3783 if (toop->is_inlinetypeptr() && !maybe_larval) {
3784 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass());
3785 res = vt;
3786 if (safe_for_replace) {
3787 replace_in_map(obj, vt);
3788 replace_in_map(not_null_obj, vt);
3789 replace_in_map(res, vt);
3790 }
3791 }
3792 }
3793 return res;
3794 }
3795
3796 Node* GraphKit::mark_word_test(Node* obj, uintptr_t mask_val, bool eq, bool check_lock) {
3797 // Load markword
3798 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3799 Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3800 if (check_lock && !UseCompactObjectHeaders) {
3801 // COH: Locking does not override the markword with a tagged pointer. We can directly read from the markword.
3802 // Check if obj is locked
3803 Node* locked_bit = MakeConX(markWord::unlocked_value);
3804 locked_bit = _gvn.transform(new AndXNode(locked_bit, mark));
3805 Node* cmp = _gvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
3806 Node* is_unlocked = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3807 IfNode* iff = new IfNode(control(), is_unlocked, PROB_MAX, COUNT_UNKNOWN);
3808 _gvn.transform(iff);
3809 Node* locked_region = new RegionNode(3);
3810 Node* mark_phi = new PhiNode(locked_region, TypeX_X);
3811
3812 // Unlocked: Use bits from mark word
3813 locked_region->init_req(1, _gvn.transform(new IfTrueNode(iff)));
3814 mark_phi->init_req(1, mark);
3815
3816 // Locked: Load prototype header from klass
3817 set_control(_gvn.transform(new IfFalseNode(iff)));
3818 // Make loads control dependent to make sure they are only executed if array is locked
3819 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
3820 Node* klass = _gvn.transform(LoadKlassNode::make(_gvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3821 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
3822 Node* proto = _gvn.transform(LoadNode::make(_gvn, control(), C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
3823
3824 locked_region->init_req(2, control());
3825 mark_phi->init_req(2, proto);
3826 set_control(_gvn.transform(locked_region));
3827 record_for_igvn(locked_region);
3828
3829 mark = mark_phi;
3830 }
3831
3832 // Now check if mark word bits are set
3833 Node* mask = MakeConX(mask_val);
3834 Node* masked = _gvn.transform(new AndXNode(_gvn.transform(mark), mask));
3835 record_for_igvn(masked); // Give it a chance to be optimized out by IGVN
3836 Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3837 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3838 }
3839
3840 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3841 return mark_word_test(obj, markWord::inline_type_pattern, is_inline, /* check_lock = */ false);
3842 }
3843
3844 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
3845 // We can't use immutable memory here because the mark word is mutable.
3846 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3847 // check is moved out of loops (mainly to enable loop unswitching).
3848 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, memory(Compile::AliasIdxRaw), array_or_klass));
3849 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3850 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3851 }
3852
3853 Node* GraphKit::null_free_array_test(Node* array, bool null_free) {
3854 return mark_word_test(array, markWord::null_free_array_bit_in_place, null_free);
3855 }
3856
3857 Node* GraphKit::null_free_atomic_array_test(Node* array, ciInlineKlass* vk) {
3858 assert(vk->has_atomic_layout() || vk->has_non_atomic_layout(), "Can't be null-free and flat");
3859
3860 // TODO 8350865 Add a stress flag to always access atomic if layout exists?
3861 if (!vk->has_non_atomic_layout()) {
3862 return intcon(1); // Always atomic
3863 } else if (!vk->has_atomic_layout()) {
3864 return intcon(0); // Never atomic
3865 }
3866
3867 Node* array_klass = load_object_klass(array);
3868 int layout_kind_offset = in_bytes(FlatArrayKlass::layout_kind_offset());
3869 Node* layout_kind_addr = basic_plus_adr(array_klass, array_klass, layout_kind_offset);
3870 Node* layout_kind = make_load(nullptr, layout_kind_addr, TypeInt::INT, T_INT, MemNode::unordered);
3871 Node* cmp = _gvn.transform(new CmpINode(layout_kind, intcon((int)LayoutKind::ATOMIC_FLAT)));
3872 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3873 }
3874
3875 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
3876 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
3877 RegionNode* region = new RegionNode(3);
3878 Node* null_ctl = top();
3879 null_check_oop(val, &null_ctl);
3880 if (null_ctl != top()) {
3881 PreserveJVMState pjvms(this);
3882 set_control(null_ctl);
3883 {
3884 // Deoptimize if null-free array
3885 BuildCutout unless(this, null_free_array_test(ary, /* null_free = */ false), PROB_MAX);
3886 inc_sp(nargs);
3887 uncommon_trap(Deoptimization::Reason_null_check,
3888 Deoptimization::Action_none);
3889 }
3890 region->init_req(1, control());
3891 }
3892 region->init_req(2, control());
3893 set_control(_gvn.transform(region));
3894 record_for_igvn(region);
3895 if (_gvn.type(val) == TypePtr::NULL_PTR) {
3896 // Since we were just successfully storing null, the array can't be null free.
3897 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
3898 ary_t = ary_t->cast_to_not_null_free();
3899 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
3900 if (safe_for_replace) {
3901 replace_in_map(ary, cast);
3902 }
3903 ary = cast;
3904 }
3905 return ary;
3906 }
3907
3908 //------------------------------next_monitor-----------------------------------
3909 // What number should be given to the next monitor?
3910 int GraphKit::next_monitor() {
3911 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3912 int next = current + C->sync_stack_slots();
3913 // Keep the toplevel high water mark current:
3914 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3915 return current;
3916 }
3917
3918 //------------------------------insert_mem_bar---------------------------------
3919 // Memory barrier to avoid floating things around
3920 // The membar serves as a pinch point between both control and all memory slices.
3921 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3922 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3923 mb->init_req(TypeFunc::Control, control());
3924 mb->init_req(TypeFunc::Memory, reset_memory());
3925 Node* membar = _gvn.transform(mb);
4017 lock->create_lock_counter(map()->jvms());
4018 increment_counter(lock->counter()->addr());
4019 }
4020 #endif
4021
4022 return flock;
4023 }
4024
4025
4026 //------------------------------shared_unlock----------------------------------
4027 // Emit unlocking code.
4028 void GraphKit::shared_unlock(Node* box, Node* obj) {
4029 // bci is either a monitorenter bc or InvocationEntryBci
4030 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
4031 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
4032
4033 if (stopped()) { // Dead monitor?
4034 map()->pop_monitor(); // Kill monitor from debug info
4035 return;
4036 }
4037 assert(!obj->is_InlineType(), "should not unlock on inline type");
4038
4039 // Memory barrier to avoid floating things down past the locked region
4040 insert_mem_bar(Op_MemBarReleaseLock);
4041
4042 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
4043 UnlockNode *unlock = new UnlockNode(C, tf);
4044 #ifdef ASSERT
4045 unlock->set_dbg_jvms(sync_jvms());
4046 #endif
4047 uint raw_idx = Compile::AliasIdxRaw;
4048 unlock->init_req( TypeFunc::Control, control() );
4049 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
4050 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
4051 unlock->init_req( TypeFunc::FramePtr, frameptr() );
4052 unlock->init_req( TypeFunc::ReturnAdr, top() );
4053
4054 unlock->init_req(TypeFunc::Parms + 0, obj);
4055 unlock->init_req(TypeFunc::Parms + 1, box);
4056 unlock = _gvn.transform(unlock)->as_Unlock();
4057
4058 Node* mem = reset_memory();
4059
4060 // unlock has no side-effects, sets few values
4061 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
4062
4063 // Kill monitor from debug info
4064 map()->pop_monitor( );
4065 }
4066
4067 //-------------------------------get_layout_helper-----------------------------
4068 // If the given klass is a constant or known to be an array,
4069 // fetch the constant layout helper value into constant_value
4070 // and return null. Otherwise, load the non-constant
4071 // layout helper value, and return the node which represents it.
4072 // This two-faced routine is useful because allocation sites
4073 // almost always feature constant types.
4074 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
4075 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
4076 if (!StressReflectiveCode && klass_t != nullptr) {
4077 bool xklass = klass_t->klass_is_exact();
4078 bool can_be_flat = false;
4079 const TypeAryPtr* ary_type = klass_t->as_instance_type()->isa_aryptr();
4080 if (UseArrayFlattening && !xklass && ary_type != nullptr && !ary_type->is_null_free()) {
4081 // Don't constant fold if the runtime type might be a flat array but the static type is not.
4082 const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
4083 can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->maybe_flat_in_array());
4084 }
4085 if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) {
4086 jint lhelper;
4087 if (klass_t->is_flat()) {
4088 lhelper = ary_type->flat_layout_helper();
4089 } else if (klass_t->isa_aryklassptr()) {
4090 BasicType elem = ary_type->elem()->array_element_basic_type();
4091 if (is_reference_type(elem, true)) {
4092 elem = T_OBJECT;
4093 }
4094 lhelper = Klass::array_layout_helper(elem);
4095 } else {
4096 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
4097 }
4098 if (lhelper != Klass::_lh_neutral_value) {
4099 constant_value = lhelper;
4100 return (Node*) nullptr;
4101 }
4102 }
4103 }
4104 constant_value = Klass::_lh_neutral_value; // put in a known value
4105 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
4106 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
4107 }
4108
4109 // We just put in an allocate/initialize with a big raw-memory effect.
4110 // Hook selected additional alias categories on the initialization.
4111 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
4112 MergeMemNode* init_in_merge,
4113 Node* init_out_raw) {
4114 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
4115 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
4116
4117 Node* prevmem = kit.memory(alias_idx);
4118 init_in_merge->set_memory_at(alias_idx, prevmem);
4119 if (init_out_raw != nullptr) {
4120 kit.set_memory(init_out_raw, alias_idx);
4121 }
4122 }
4123
4124 //---------------------------set_output_for_allocation-------------------------
4125 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
4126 const TypeOopPtr* oop_type,
4127 bool deoptimize_on_exception) {
4128 int rawidx = Compile::AliasIdxRaw;
4129 alloc->set_req( TypeFunc::FramePtr, frameptr() );
4130 add_safepoint_edges(alloc);
4131 Node* allocx = _gvn.transform(alloc);
4132 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
4133 // create memory projection for i_o
4134 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
4135 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
4136
4137 // create a memory projection as for the normal control path
4138 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
4139 set_memory(malloc, rawidx);
4140
4141 // a normal slow-call doesn't change i_o, but an allocation does
4142 // we create a separate i_o projection for the normal control path
4143 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
4144 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
4145
4146 // put in an initialization barrier
4147 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
4148 rawoop)->as_Initialize();
4149 assert(alloc->initialization() == init, "2-way macro link must work");
4150 assert(init ->allocation() == alloc, "2-way macro link must work");
4151 {
4152 // Extract memory strands which may participate in the new object's
4153 // initialization, and source them from the new InitializeNode.
4154 // This will allow us to observe initializations when they occur,
4155 // and link them properly (as a group) to the InitializeNode.
4156 assert(init->in(InitializeNode::Memory) == malloc, "");
4157 MergeMemNode* minit_in = MergeMemNode::make(malloc);
4158 init->set_req(InitializeNode::Memory, minit_in);
4159 record_for_igvn(minit_in); // fold it up later, if possible
4160 _gvn.set_type(minit_in, Type::MEMORY);
4161 Node* minit_out = memory(rawidx);
4162 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
4163 // Add an edge in the MergeMem for the header fields so an access
4164 // to one of those has correct memory state
4165 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
4166 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
4167 if (oop_type->isa_aryptr()) {
4168 const TypeAryPtr* arytype = oop_type->is_aryptr();
4169 if (arytype->is_flat()) {
4170 // Initially all flat array accesses share a single slice
4171 // but that changes after parsing. Prepare the memory graph so
4172 // it can optimize flat array accesses properly once they
4173 // don't share a single slice.
4174 assert(C->flat_accesses_share_alias(), "should be set at parse time");
4175 C->set_flat_accesses_share_alias(false);
4176 ciInlineKlass* vk = arytype->elem()->inline_klass();
4177 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) {
4178 ciField* field = vk->nonstatic_field_at(i);
4179 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4180 continue; // do not bother to track really large numbers of fields
4181 int off_in_vt = field->offset_in_bytes() - vk->payload_offset();
4182 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
4183 int fieldidx = C->get_alias_index(adr_type, true);
4184 // Pass nullptr for init_out. Having per flat array element field memory edges as uses of the Initialize node
4185 // can result in per flat array field Phis to be created which confuses the logic of
4186 // Compile::adjust_flat_array_access_aliases().
4187 hook_memory_on_init(*this, fieldidx, minit_in, nullptr);
4188 }
4189 C->set_flat_accesses_share_alias(true);
4190 hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::INLINES), minit_in, minit_out);
4191 } else {
4192 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
4193 int elemidx = C->get_alias_index(telemref);
4194 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
4195 }
4196 } else if (oop_type->isa_instptr()) {
4197 set_memory(minit_out, C->get_alias_index(oop_type)); // mark word
4198 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
4199 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
4200 ciField* field = ik->nonstatic_field_at(i);
4201 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4202 continue; // do not bother to track really large numbers of fields
4203 // Find (or create) the alias category for this field:
4204 int fieldidx = C->alias_type(field)->index();
4205 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
4206 }
4207 }
4208 }
4209
4210 // Cast raw oop to the real thing...
4211 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
4212 javaoop = _gvn.transform(javaoop);
4213 C->set_recent_alloc(control(), javaoop);
4214 assert(just_allocated_object(control()) == javaoop, "just allocated");
4215
4216 #ifdef ASSERT
4217 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
4228 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
4229 }
4230 }
4231 #endif //ASSERT
4232
4233 return javaoop;
4234 }
4235
4236 //---------------------------new_instance--------------------------------------
4237 // This routine takes a klass_node which may be constant (for a static type)
4238 // or may be non-constant (for reflective code). It will work equally well
4239 // for either, and the graph will fold nicely if the optimizer later reduces
4240 // the type to a constant.
4241 // The optional arguments are for specialized use by intrinsics:
4242 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
4243 // - If 'return_size_val', report the total object size to the caller.
4244 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4245 Node* GraphKit::new_instance(Node* klass_node,
4246 Node* extra_slow_test,
4247 Node* *return_size_val,
4248 bool deoptimize_on_exception,
4249 InlineTypeNode* inline_type_node) {
4250 // Compute size in doublewords
4251 // The size is always an integral number of doublewords, represented
4252 // as a positive bytewise size stored in the klass's layout_helper.
4253 // The layout_helper also encodes (in a low bit) the need for a slow path.
4254 jint layout_con = Klass::_lh_neutral_value;
4255 Node* layout_val = get_layout_helper(klass_node, layout_con);
4256 bool layout_is_con = (layout_val == nullptr);
4257
4258 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
4259 // Generate the initial go-slow test. It's either ALWAYS (return a
4260 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
4261 // case) a computed value derived from the layout_helper.
4262 Node* initial_slow_test = nullptr;
4263 if (layout_is_con) {
4264 assert(!StressReflectiveCode, "stress mode does not use these paths");
4265 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4266 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4267 } else { // reflective case
4268 // This reflective path is used by Unsafe.allocateInstance.
4269 // (It may be stress-tested by specifying StressReflectiveCode.)
4270 // Basically, we want to get into the VM is there's an illegal argument.
4271 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4272 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4273 if (extra_slow_test != intcon(0)) {
4274 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4275 }
4276 // (Macro-expander will further convert this to a Bool, if necessary.)
4287
4288 // Clear the low bits to extract layout_helper_size_in_bytes:
4289 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4290 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4291 size = _gvn.transform( new AndXNode(size, mask) );
4292 }
4293 if (return_size_val != nullptr) {
4294 (*return_size_val) = size;
4295 }
4296
4297 // This is a precise notnull oop of the klass.
4298 // (Actually, it need not be precise if this is a reflective allocation.)
4299 // It's what we cast the result to.
4300 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4301 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
4302 const TypeOopPtr* oop_type = tklass->as_instance_type();
4303
4304 // Now generate allocation code
4305
4306 // The entire memory state is needed for slow path of the allocation
4307 // since GC and deoptimization can happen.
4308 Node *mem = reset_memory();
4309 set_all_memory(mem); // Create new memory state
4310
4311 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4312 control(), mem, i_o(),
4313 size, klass_node,
4314 initial_slow_test, inline_type_node);
4315
4316 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4317 }
4318
4319 //-------------------------------new_array-------------------------------------
4320 // helper for newarray and anewarray
4321 // The 'length' parameter is (obviously) the length of the array.
4322 // The optional arguments are for specialized use by intrinsics:
4323 // - If 'return_size_val', report the non-padded array size (sum of header size
4324 // and array body) to the caller.
4325 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4326 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4327 Node* length, // number of array elements
4328 int nargs, // number of arguments to push back for uncommon trap
4329 Node* *return_size_val,
4330 bool deoptimize_on_exception,
4331 Node* init_val) {
4332 jint layout_con = Klass::_lh_neutral_value;
4333 Node* layout_val = get_layout_helper(klass_node, layout_con);
4334 bool layout_is_con = (layout_val == nullptr);
4335
4336 if (!layout_is_con && !StressReflectiveCode &&
4337 !too_many_traps(Deoptimization::Reason_class_check)) {
4338 // This is a reflective array creation site.
4339 // Optimistically assume that it is a subtype of Object[],
4340 // so that we can fold up all the address arithmetic.
4341 layout_con = Klass::array_layout_helper(T_OBJECT);
4342 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4343 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4344 { BuildCutout unless(this, bol_lh, PROB_MAX);
4345 inc_sp(nargs);
4346 uncommon_trap(Deoptimization::Reason_class_check,
4347 Deoptimization::Action_maybe_recompile);
4348 }
4349 layout_val = nullptr;
4350 layout_is_con = true;
4351 }
4352
4353 // Generate the initial go-slow test. Make sure we do not overflow
4354 // if length is huge (near 2Gig) or negative! We do not need
4355 // exact double-words here, just a close approximation of needed
4356 // double-words. We can't add any offset or rounding bits, lest we
4357 // take a size -1 of bytes and make it positive. Use an unsigned
4358 // compare, so negative sizes look hugely positive.
4359 int fast_size_limit = FastAllocateSizeLimit;
4360 if (layout_is_con) {
4361 assert(!StressReflectiveCode, "stress mode does not use these paths");
4362 // Increase the size limit if we have exact knowledge of array type.
4363 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4364 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4365 }
4366
4367 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4368 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4369
4370 // --- Size Computation ---
4371 // array_size = round_to_heap(array_header + (length << elem_shift));
4372 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4373 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4374 // The rounding mask is strength-reduced, if possible.
4375 int round_mask = MinObjAlignmentInBytes - 1;
4376 Node* header_size = nullptr;
4377 // (T_BYTE has the weakest alignment and size restrictions...)
4378 if (layout_is_con) {
4379 int hsize = Klass::layout_helper_header_size(layout_con);
4380 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4381 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4382 if ((round_mask & ~right_n_bits(eshift)) == 0)
4383 round_mask = 0; // strength-reduce it if it goes away completely
4384 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4385 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4386 assert(header_size_min <= hsize, "generic minimum is smallest");
4387 header_size = intcon(hsize);
4388 } else {
4389 Node* hss = intcon(Klass::_lh_header_size_shift);
4390 Node* hsm = intcon(Klass::_lh_header_size_mask);
4391 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
4392 header_size = _gvn.transform(new AndINode(header_size, hsm));
4393 }
4394
4395 Node* elem_shift = nullptr;
4396 if (layout_is_con) {
4397 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4398 if (eshift != 0)
4399 elem_shift = intcon(eshift);
4400 } else {
4401 // There is no need to mask or shift this value.
4402 // The semantics of LShiftINode include an implicit mask to 0x1F.
4403 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4404 elem_shift = layout_val;
4453 }
4454 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
4455
4456 if (return_size_val != nullptr) {
4457 // This is the size
4458 (*return_size_val) = non_rounded_size;
4459 }
4460
4461 Node* size = non_rounded_size;
4462 if (round_mask != 0) {
4463 Node* mask1 = MakeConX(round_mask);
4464 size = _gvn.transform(new AddXNode(size, mask1));
4465 Node* mask2 = MakeConX(~round_mask);
4466 size = _gvn.transform(new AndXNode(size, mask2));
4467 }
4468 // else if round_mask == 0, the size computation is self-rounding
4469
4470 // Now generate allocation code
4471
4472 // The entire memory state is needed for slow path of the allocation
4473 // since GC and deoptimization can happen.
4474 Node *mem = reset_memory();
4475 set_all_memory(mem); // Create new memory state
4476
4477 if (initial_slow_test->is_Bool()) {
4478 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4479 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4480 }
4481
4482 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4483 const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4484
4485 Node* raw_init_value = nullptr;
4486 if (init_val != nullptr) {
4487 // TODO 8350865 Fast non-zero init not implemented yet for flat, null-free arrays
4488 if (ary_type->is_flat()) {
4489 initial_slow_test = intcon(1);
4490 }
4491
4492 if (UseCompressedOops) {
4493 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4494 init_val = _gvn.transform(new EncodePNode(init_val, init_val->bottom_type()->make_narrowoop()));
4495 Node* lower = _gvn.transform(new CastP2XNode(control(), init_val));
4496 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4497 raw_init_value = _gvn.transform(new OrLNode(lower, upper));
4498 } else {
4499 raw_init_value = _gvn.transform(new CastP2XNode(control(), init_val));
4500 }
4501 }
4502
4503 Node* valid_length_test = _gvn.intcon(1);
4504 if (ary_type->isa_aryptr()) {
4505 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4506 jint max = TypeAryPtr::max_array_length(bt);
4507 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
4508 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4509 }
4510
4511 // Create the AllocateArrayNode and its result projections
4512 AllocateArrayNode* alloc
4513 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4514 control(), mem, i_o(),
4515 size, klass_node,
4516 initial_slow_test,
4517 length, valid_length_test,
4518 init_val, raw_init_value);
4519 // Cast to correct type. Note that the klass_node may be constant or not,
4520 // and in the latter case the actual array type will be inexact also.
4521 // (This happens via a non-constant argument to inline_native_newArray.)
4522 // In any case, the value of klass_node provides the desired array type.
4523 const TypeInt* length_type = _gvn.find_int_type(length);
4524 if (ary_type->isa_aryptr() && length_type != nullptr) {
4525 // Try to get a better type than POS for the size
4526 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4527 }
4528
4529 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4530
4531 array_ideal_length(alloc, ary_type, true);
4532 return javaoop;
4533 }
4534
4535 // The following "Ideal_foo" functions are placed here because they recognize
4536 // the graph shapes created by the functions immediately above.
4537
4538 //---------------------------Ideal_allocation----------------------------------
4633 void GraphKit::add_parse_predicates(int nargs) {
4634 if (ShortRunningLongLoop) {
4635 // Will narrow the limit down with a cast node. Predicates added later may depend on the cast so should be last when
4636 // walking up from the loop.
4637 add_parse_predicate(Deoptimization::Reason_short_running_long_loop, nargs);
4638 }
4639 if (UseLoopPredicate) {
4640 add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4641 if (UseProfiledLoopPredicate) {
4642 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4643 }
4644 }
4645 if (UseAutoVectorizationPredicate) {
4646 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, nargs);
4647 }
4648 // Loop Limit Check Predicate should be near the loop.
4649 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4650 }
4651
4652 void GraphKit::sync_kit(IdealKit& ideal) {
4653 reset_memory();
4654 set_all_memory(ideal.merged_memory());
4655 set_i_o(ideal.i_o());
4656 set_control(ideal.ctrl());
4657 }
4658
4659 void GraphKit::final_sync(IdealKit& ideal) {
4660 // Final sync IdealKit and graphKit.
4661 sync_kit(ideal);
4662 }
4663
4664 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4665 Node* len = load_array_length(load_String_value(str, set_ctrl));
4666 Node* coder = load_String_coder(str, set_ctrl);
4667 // Divide length by 2 if coder is UTF16
4668 return _gvn.transform(new RShiftINode(len, coder));
4669 }
4670
4671 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4672 int value_offset = java_lang_String::value_offset();
4673 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4674 false, nullptr, Type::Offset(0));
4675 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4676 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4677 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true),
4678 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4679 Node* p = basic_plus_adr(str, str, value_offset);
4680 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4681 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4682 return load;
4683 }
4684
4685 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4686 if (!CompactStrings) {
4687 return intcon(java_lang_String::CODER_UTF16);
4688 }
4689 int coder_offset = java_lang_String::coder_offset();
4690 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4691 false, nullptr, Type::Offset(0));
4692 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4693
4694 Node* p = basic_plus_adr(str, str, coder_offset);
4695 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4696 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4697 return load;
4698 }
4699
4700 void GraphKit::store_String_value(Node* str, Node* value) {
4701 int value_offset = java_lang_String::value_offset();
4702 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4703 false, nullptr, Type::Offset(0));
4704 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4705
4706 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4707 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4708 }
4709
4710 void GraphKit::store_String_coder(Node* str, Node* value) {
4711 int coder_offset = java_lang_String::coder_offset();
4712 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4713 false, nullptr, Type::Offset(0));
4714 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4715
4716 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4717 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4718 }
4719
4720 // Capture src and dst memory state with a MergeMemNode
4721 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4722 if (src_type == dst_type) {
4723 // Types are equal, we don't need a MergeMemNode
4724 return memory(src_type);
4725 }
4726 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4727 record_for_igvn(merge); // fold it up later, if possible
4728 int src_idx = C->get_alias_index(src_type);
4729 int dst_idx = C->get_alias_index(dst_type);
4730 merge->set_memory_at(src_idx, memory(src_idx));
4731 merge->set_memory_at(dst_idx, memory(dst_idx));
4732 return merge;
4733 }
4806 i_char->init_req(2, AddI(i_char, intcon(2)));
4807
4808 set_control(IfFalse(iff));
4809 set_memory(st, TypeAryPtr::BYTES);
4810 }
4811
4812 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4813 if (!field->is_constant()) {
4814 return nullptr; // Field not marked as constant.
4815 }
4816 ciInstance* holder = nullptr;
4817 if (!field->is_static()) {
4818 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4819 if (const_oop != nullptr && const_oop->is_instance()) {
4820 holder = const_oop->as_instance();
4821 }
4822 }
4823 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4824 /*is_unsigned_load=*/false);
4825 if (con_type != nullptr) {
4826 Node* con = makecon(con_type);
4827 if (field->type()->is_inlinetype()) {
4828 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass());
4829 } else if (con_type->is_inlinetypeptr()) {
4830 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass());
4831 }
4832 return con;
4833 }
4834 return nullptr;
4835 }
4836
4837 //---------------------------load_mirror_from_klass----------------------------
4838 // Given a klass oop, load its java mirror (a java.lang.Class oop).
4839 Node* GraphKit::load_mirror_from_klass(Node* klass) {
4840 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
4841 Node* load = make_load(nullptr, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4842 // mirror = ((OopHandle)mirror)->resolve();
4843 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
4844 }
4845
4846 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4847 const Type* obj_type = obj->bottom_type();
4848 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4849 if (obj_type->isa_oopptr() && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4850 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4851 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4852 obj = casted_obj;
4853 }
4854 if (sig_type->is_inlinetypeptr()) {
4855 obj = InlineTypeNode::make_from_oop(this, obj, sig_type->inline_klass());
4856 }
4857 return obj;
4858 }
|